Skip to content

DanLing

danling

AverageMeter

A lightweight utility to compute and store running averages of values.

AverageMeter provides an efficient way to track running statistics (current value, sum, count, average) with minimal memory overhead and support for distributed environments.

Attributes:

Name Type Description
val float

Most recent value added to the meter
bat float

Most recent value, synchronized across distributed processes
avg float

Running average of all values, weighted by counts
sum float

Sum of all values added to the meter
count int

Total count of values added (considering weights)

Examples:

Python Console Session
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
>>> meter = AverageMeter()
>>> meter.update(0.7)
>>> meter.val
0.7
>>> meter.bat  # Same as val in non-distributed settings
0.7
>>> meter.avg
0.7
>>> meter.update(0.9)
>>> meter.val
0.9
>>> meter.avg
0.8
>>> meter.sum
1.6
>>> meter.count
2
>>> # Weighted update
>>> meter.update(value=0.5, n=3)
>>> meter.avg
0.62
>>> meter.reset()
AverageMeter(val=0.0, avg=nan)
See Also
  • MetricMeter: Memory-efficient metric tracker that averages metrics batch-by-batch.
Source code in danling/metric/average_meter.py
Python
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
class AverageMeter:
    r"""
    A lightweight utility to compute and store running averages of values.

    AverageMeter provides an efficient way to track running statistics (current value, sum, count, average)
    with minimal memory overhead and support for distributed environments.

    Attributes:
        val: Most recent value added to the meter
        bat: Most recent value, synchronized across distributed processes
        avg: Running average of all values, weighted by counts
        sum: Sum of all values added to the meter
        count: Total count of values added (considering weights)

    Examples:
        >>> meter = AverageMeter()
        >>> meter.update(0.7)
        >>> meter.val
        0.7
        >>> meter.bat  # Same as val in non-distributed settings
        0.7
        >>> meter.avg
        0.7
        >>> meter.update(0.9)
        >>> meter.val
        0.9
        >>> meter.avg
        0.8
        >>> meter.sum
        1.6
        >>> meter.count
        2
        >>> # Weighted update
        >>> meter.update(value=0.5, n=3)
        >>> meter.avg
        0.62
        >>> meter.reset()
        AverageMeter(val=0.0, avg=nan)

    See Also:
        - [`MetricMeter`][danling.metric.metric_meter.MetricMeter]:
            Memory-efficient metric tracker that averages metrics batch-by-batch.
    """

    v: float = 0.0
    n: int = 1
    sum: float = 0.0
    count: int = 0

    def __init__(self) -> None:
        self.reset()

    def reset(self) -> Self:
        r"""
        Resets the meter.
        """

        self.v = 0.0
        self.n = 1
        self.sum = 0.0
        self.count = 0
        return self

    def update(self, value: float | int, n: int = 1) -> None:
        r"""
        Updates the average and current value in the meter.

        Args:
            value: Value to be added to the average.
            n: Number of values to be added.
        """

        self.v = value
        self.n = n
        self.sum += value * n
        self.count += n

    def value(self) -> float:
        return self.v

    def batch(self) -> float:
        world_size = get_world_size()
        if world_size <= 1:
            return self.value()
        synced_tensor = torch.tensor([self.v, self.n], dtype=torch.float64).cuda()
        dist.all_reduce(synced_tensor)
        val, count = synced_tensor
        if count == 0:
            return float("nan")
        return (val / count).item()

    def average(self) -> float:
        world_size = get_world_size()
        if world_size <= 1:
            return self.sum / self.count if self.count != 0 else float("nan")
        synced_tensor = torch.tensor([self.sum, self.count], dtype=torch.float64).cuda()
        dist.all_reduce(synced_tensor)
        val, count = synced_tensor
        if count == 0:
            return float("nan")
        return (val / count).item()

    @property
    def val(self) -> float:
        return self.value()

    @property
    def bat(self) -> float:
        return self.batch()

    @property
    def avg(self) -> float:
        return self.average()

    def __format__(self, format_spec: str) -> str:
        return f"{self.val.__format__(format_spec)} ({self.avg.__format__(format_spec)})"

    def __repr__(self) -> str:
        return f"{self.__class__.__name__}(val={self.val}, avg={self.avg})"

reset

Python
reset() -> Self

Resets the meter.

Source code in danling/metric/average_meter.py
Python
 92
 93
 94
 95
 96
 97
 98
 99
100
101
def reset(self) -> Self:
    r"""
    Resets the meter.
    """

    self.v = 0.0
    self.n = 1
    self.sum = 0.0
    self.count = 0
    return self

update

Python
update(value: float | int, n: int = 1) -> None

Updates the average and current value in the meter.

Parameters:

Name Type Description Default
value
float | int

Value to be added to the average.

 required
n
int

Number of values to be added.

 1
Source code in danling/metric/average_meter.py
Python
103
104
105
106
107
108
109
110
111
112
113
114
115
def update(self, value: float | int, n: int = 1) -> None:
    r"""
    Updates the average and current value in the meter.

    Args:
        value: Value to be added to the average.
        n: Number of values to be added.
    """

    self.v = value
    self.n = n
    self.sum += value * n
    self.count += n

AverageMeters

Bases: MetricsDict

Manages multiple average meters in one object.

Examples:

Python Console Session
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
>>> meters = AverageMeters()
>>> meters.update({"loss": 0.6, "auroc": 0.7, "r2": 0.8})
>>> f"{meters:.4f}"
'loss: 0.6000 (0.6000)\tauroc: 0.7000 (0.7000)\tr2: 0.8000 (0.8000)'
>>> meters['loss'].update(value=0.9, n=1)
>>> f"{meters:.4f}"
'loss: 0.9000 (0.7500)\tauroc: 0.7000 (0.7000)\tr2: 0.8000 (0.8000)'
>>> meters.sum.dict()
{'loss': 1.5, 'auroc': 0.7, 'r2': 0.8}
>>> meters.count.dict()
{'loss': 2, 'auroc': 1, 'r2': 1}
>>> meters.reset()
AverageMeters(...)
>>> f"{meters:.4f}"
'loss: 0.0000 (nan)\tauroc: 0.0000 (nan)\tr2: 0.0000 (nan)'
See Also
  • MetricMeters: Memory-efficient metric tracker that averages multiple metrics batch-by-batch.
Source code in danling/metric/average_meter.py
Python
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
class AverageMeters(MetricsDict):
    r"""
    Manages multiple average meters in one object.

    Examples:
        >>> meters = AverageMeters()
        >>> meters.update({"loss": 0.6, "auroc": 0.7, "r2": 0.8})
        >>> f"{meters:.4f}"
        'loss: 0.6000 (0.6000)\tauroc: 0.7000 (0.7000)\tr2: 0.8000 (0.8000)'
        >>> meters['loss'].update(value=0.9, n=1)
        >>> f"{meters:.4f}"
        'loss: 0.9000 (0.7500)\tauroc: 0.7000 (0.7000)\tr2: 0.8000 (0.8000)'
        >>> meters.sum.dict()
        {'loss': 1.5, 'auroc': 0.7, 'r2': 0.8}
        >>> meters.count.dict()
        {'loss': 2, 'auroc': 1, 'r2': 1}
        >>> meters.reset()
        AverageMeters(...)
        >>> f"{meters:.4f}"
        'loss: 0.0000 (nan)\tauroc: 0.0000 (nan)\tr2: 0.0000 (nan)'

    See Also:
        - [`MetricMeters`][danling.metric.metric_meter.MetricMeters]:
            Memory-efficient metric tracker that averages multiple metrics batch-by-batch.
    """

    def __init__(self, default_factory: Type[AverageMeter] = AverageMeter, **meters) -> None:
        for name, meter in meters.items():
            if not isinstance(meter, AverageMeter):
                raise ValueError(f"Expected {name} to be an instance of AverageMeter, but got {type(meter)}")
        super().__init__(default_factory=default_factory, **meters)

    @property
    def sum(self) -> FlatDict[str, float]:
        return FlatDict({key: meter.sum for key, meter in self.all_items()})

    @property
    def count(self) -> FlatDict[str, int]:
        return FlatDict({key: meter.count for key, meter in self.all_items()})

    def update(self, *args: Dict, **values: int | float) -> None:  # pylint: disable=W0237
        r"""
        Updates the average and current value in all meters.

        Args:
            values: Dict of values to be added to the average.
            n: Number of values to be added.

        Raises:
            ValueError: If the value is not an instance of (int, float).
        """  # noqa: E501

        if args:
            if len(args) > 1:
                raise ValueError("Expected only one positional argument, but got multiple.")
            values = args[0].update(values) or args[0] if values else args[0]

        for meter, value in values.items():
            if isinstance(value, Tensor):
                value = value.item()
            if not isinstance(value, (int, float)):
                raise ValueError(f"Expected values to be int or float, but got {type(value)}")
            self[meter].update(value)

    def set(self, name: str, meter: AverageMeter) -> None:  # pylint: disable=W0237
        if not isinstance(meter, AverageMeter):
            raise ValueError(f"Expected meter to be an instance of AverageMeter, but got {type(meter)}")
        super().set(name, meter)

update

Python
update(*args: Dict, **values: int | float) -> None

Updates the average and current value in all meters.

Parameters:

Name Type Description Default
values
int | float

Dict of values to be added to the average.

 {}
n

Number of values to be added.

 required

Raises:

Type Description
ValueError

If the value is not an instance of (int, float).
Source code in danling/metric/average_meter.py
Python
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
def update(self, *args: Dict, **values: int | float) -> None:  # pylint: disable=W0237
    r"""
    Updates the average and current value in all meters.

    Args:
        values: Dict of values to be added to the average.
        n: Number of values to be added.

    Raises:
        ValueError: If the value is not an instance of (int, float).
    """  # noqa: E501

    if args:
        if len(args) > 1:
            raise ValueError("Expected only one positional argument, but got multiple.")
        values = args[0].update(values) or args[0] if values else args[0]

    for meter, value in values.items():
        if isinstance(value, Tensor):
            value = value.item()
        if not isinstance(value, (int, float)):
            raise ValueError(f"Expected values to be int or float, but got {type(value)}")
        self[meter].update(value)

MetricMeter

Bases: AverageMeter

A memory-efficient metric tracker that computes and averages metrics across batches.

MetricMeter applies a metric function to each batch and maintains running averages without storing the complete history of predictions and labels. This makes it ideal for metrics that can be meaningfully averaged across batches (like accuracy or loss).

Attributes:

Name Type Description
metric Callable

The metric function to compute on each batch
preprocess Optional[Callable]

Optional preprocessing function to apply to inputs and targets
ignore_index int

Value to ignore in classification tasks (e.g., -100 for padding)
ignore_nan bool

Whether to ignore NaN values in regression tasks
val float

Result from the most recent batch
bat float

Result from the most recent batch, synchronized across devices
avg float

Weighted average of all results so far
sum float

Running sum of (metric × batch_size) values
count int

Running sum of batch sizes

Parameters:

Name Type Description Default

metric

Callable

Function that computes a metric given input and target tensors

 required

preprocess

Callable | None

Function to preprocess inputs before computing the metric

 None

ignore_index

int | None

Value to ignore in classification tasks

 None

ignore_nan

bool | None

Whether to ignore NaN values in regression tasks

 None

Examples:

Python Console Session
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
>>> from danling.metric.functional import accuracy
>>> meter = MetricMeter(accuracy)
>>> meter.update([0.1, 0.8, 0.6, 0.2], [0, 1, 0, 0])
>>> meter.val
0.75
>>> meter.avg
0.75
>>> meter.update([0.1, 0.7, 0.3, 0.2, 0.8, 0.4], [0, 1, 1, 0, 0, 1])
>>> meter.val
0.5
>>> meter.avg
0.6
>>> meter.sum
6.0
>>> meter.count
10
>>> meter.reset()
MetricMeter(accuracy)
>>> meter.val
0.0
>>> meter.avg
nan
Notes
  • MetricMeter is more memory-efficient than Metrics because it only stores running statistics
  • Only suitable for metrics that can be meaningfully averaged batch-by-batch
  • Not suitable for metrics like AUROC that need the entire dataset
  • The metric function should accept input and target tensors and return a scalar value
  • For multiple metrics, use MetricMeters
See Also
  • AverageMeter: A lightweight utility to compute and store running averages of values.
Source code in danling/metric/metric_meter.py
Python
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
class MetricMeter(AverageMeter):
    r"""
    A memory-efficient metric tracker that computes and averages metrics across batches.

    MetricMeter applies a metric function to each batch and maintains running averages
    without storing the complete history of predictions and labels. This makes it ideal for
    metrics that can be meaningfully averaged across batches (like accuracy or loss).

    Attributes:
        metric: The metric function to compute on each batch
        preprocess: Optional preprocessing function to apply to inputs and targets
        ignore_index: Value to ignore in classification tasks (e.g., -100 for padding)
        ignore_nan: Whether to ignore NaN values in regression tasks
        val: Result from the most recent batch
        bat: Result from the most recent batch, synchronized across devices
        avg: Weighted average of all results so far
        sum: Running sum of (metric × batch_size) values
        count: Running sum of batch sizes

    Args:
        metric: Function that computes a metric given input and target tensors
        preprocess: Function to preprocess inputs before computing the metric
        ignore_index: Value to ignore in classification tasks
        ignore_nan: Whether to ignore NaN values in regression tasks

    Examples:
        >>> from danling.metric.functional import accuracy
        >>> meter = MetricMeter(accuracy)
        >>> meter.update([0.1, 0.8, 0.6, 0.2], [0, 1, 0, 0])
        >>> meter.val
        0.75
        >>> meter.avg
        0.75
        >>> meter.update([0.1, 0.7, 0.3, 0.2, 0.8, 0.4], [0, 1, 1, 0, 0, 1])
        >>> meter.val
        0.5
        >>> meter.avg
        0.6
        >>> meter.sum
        6.0
        >>> meter.count
        10
        >>> meter.reset()
        MetricMeter(accuracy)
        >>> meter.val
        0.0
        >>> meter.avg
        nan

    Notes:
        - MetricMeter is more memory-efficient than [`Metrics`][danling.metric.metrics.Metrics]
          because it only stores running statistics
        - Only suitable for metrics that can be meaningfully averaged batch-by-batch
        - Not suitable for metrics like AUROC that need the entire dataset
        - The metric function should accept input and target tensors and return a scalar value
        - For multiple metrics, use [`MetricMeters`][danling.metric.metric_meter.MetricMeters]

    See Also:
        - [`AverageMeter`][danling.metric.average_meter.AverageMeter]:
            A lightweight utility to compute and store running averages of values.
    """

    metric: Callable
    preprocess: Optional[Callable] = None
    ignore_index: int = -100
    ignore_nan: bool = True

    def __init__(
        self,
        metric: Callable,
        preprocess: Callable | None = None,
        ignore_index: int | None = None,
        ignore_nan: bool | None = None,
    ) -> None:
        super().__init__()
        self.preprocess = preprocess
        if ignore_index is not None:
            self.ignore_index = ignore_index
        if ignore_nan is not None:
            self.ignore_nan = ignore_nan
        if preprocess is not None and "preprocess" in signature(metric).parameters:
            metric = partial(metric, preprocess=None)
        self.metric = metric

    def update(  # type: ignore[override] # pylint: disable=W0237
        self,
        input: Tensor | NestedTensor | Sequence,  # pylint: disable=W0622
        target: Tensor | NestedTensor | Sequence,
        preprocessed: bool = False,
    ) -> None:
        r"""
        Updates the average and current value in the meter.

        Args:
            value: Value to be added to the average.
            n: Number of values to be added.
        """
        if self.preprocess is not None and not preprocessed:
            input, target = self.preprocess(input, target, ignore_index=self.ignore_index, ignore_nan=self.ignore_nan)
        n = len(input)
        value = self.metric(input, target)
        if isinstance(value, Tensor):
            value = value.item()
        super().update(value=value, n=n)

    def __repr__(self):
        metric = self.metric
        if isinstance(metric, partial):
            metric = metric.func
        return f"{self.__class__.__name__}({metric.__name__})"

update

Python
update(
    input: Tensor | NestedTensor | Sequence,
    target: Tensor | NestedTensor | Sequence,
    preprocessed: bool = False,
) -> None

Updates the average and current value in the meter.

Parameters:

Name Type Description Default
value

Value to be added to the average.

 required
n

Number of values to be added.

 required
Source code in danling/metric/metric_meter.py
Python
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
def update(  # type: ignore[override] # pylint: disable=W0237
    self,
    input: Tensor | NestedTensor | Sequence,  # pylint: disable=W0622
    target: Tensor | NestedTensor | Sequence,
    preprocessed: bool = False,
) -> None:
    r"""
    Updates the average and current value in the meter.

    Args:
        value: Value to be added to the average.
        n: Number of values to be added.
    """
    if self.preprocess is not None and not preprocessed:
        input, target = self.preprocess(input, target, ignore_index=self.ignore_index, ignore_nan=self.ignore_nan)
    n = len(input)
    value = self.metric(input, target)
    if isinstance(value, Tensor):
        value = value.item()
    super().update(value=value, n=n)

MetricMeters

Bases: AverageMeters

A container for managing multiple MetricMeter instances with shared preprocessing.

MetricMeters allows you to organize and track multiple metrics in a unified interface, with consistent preprocessing applied to all inputs before computing each metric. This is particularly useful when you want to track several metrics that can be meaningfully averaged across batches.

Attributes:

Name Type Description
preprocess

Shared preprocessing function for all meters
ignore_index

Value to ignore in classification tasks
ignore_nan

Whether to ignore NaN values in regression tasks
val NestedDict[str, float]

Dictionary of current values from all meters
avg NestedDict[str, float]

Dictionary of running averages from all meters
sum FlatDict[str, float]

Dictionary of sums from all meters
count FlatDict[str, int]

Dictionary of counts from all meters

Parameters:

Name Type Description Default

*args

Either metric functions or a Metrics instance to extract metrics from

 ()

preprocess

Callable | None

Preprocessing function to apply to inputs before computing metrics

 None

ignore_index

int | None

Value to ignore in classification tasks

 None

ignore_nan

bool | None

Whether to ignore NaN values in regression tasks

 None

**meters

Named MetricMeter instances or metric functions

 {}

Examples:

Python Console Session
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
>>> from danling.metric.functional import accuracy, auroc, auprc, base_preprocess
>>> meters = MetricMeters(acc=accuracy, auroc=auroc, auprc=auprc, preprocess=base_preprocess)
>>> meters.update([0.1, 0.8, 0.6, 0.2], [0, 1, 0, 0])
>>> meters.sum.dict()
{'acc': 3.0, 'auroc': 4.0, 'auprc': 4.0}
>>> meters.count.dict()
{'acc': 4, 'auroc': 4, 'auprc': 4}
>>> meters['auroc'].update([0.2, 0.8], [0, 1])
>>> meters.sum.dict()
{'acc': 3.0, 'auroc': 6.0, 'auprc': 4.0}
>>> meters.count.dict()
{'acc': 4, 'auroc': 6, 'auprc': 4}
>>> meters.update([[0.1, 0.7, 0.3, 0.2], [0.8, 0.4]], [[0, 0, 1, 0], [0, 0]])
>>> meters.sum.dict()
{'acc': 6.0, 'auroc': 8.4, 'auprc': 5.5}
>>> meters.count.dict()
{'acc': 10, 'auroc': 12, 'auprc': 10}
>>> meters['auroc'].update([0.4, 0.8, 0.6, 0.2], [0, 1, 1, 0])
>>> meters.avg.dict()
{'acc': 0.6, 'auroc': 0.775, 'auprc': 0.55}
>>> meters.update(dict(loss=""))
Traceback (most recent call last):
TypeError: ...update() missing 1 required positional argument: 'target'
Notes
  • MetricMeters manages multiple MetricMeter instances with shared preprocessing
  • Each metric is computed independently but uses the same inputs
  • All meters are updated simultaneously when you call update()
  • Individual meters can be accessed like dictionary items or attributes
See Also
  • AverageMeters: A container for managing multiple average meters in one object.
  • Metrics: Metric tracker that stores the complete prediction and target history.
Source code in danling/metric/metric_meter.py
Python
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
class MetricMeters(AverageMeters):
    r"""
    A container for managing multiple MetricMeter instances with shared preprocessing.

    MetricMeters allows you to organize and track multiple metrics in a unified interface,
    with consistent preprocessing applied to all inputs before computing each metric.
    This is particularly useful when you want to track several metrics that can be
    meaningfully averaged across batches.

    Attributes:
        preprocess: Shared preprocessing function for all meters
        ignore_index: Value to ignore in classification tasks
        ignore_nan: Whether to ignore NaN values in regression tasks
        val: Dictionary of current values from all meters
        avg: Dictionary of running averages from all meters
        sum: Dictionary of sums from all meters
        count: Dictionary of counts from all meters

    Args:
        *args: Either metric functions or a Metrics instance to extract metrics from
        preprocess: Preprocessing function to apply to inputs before computing metrics
        ignore_index: Value to ignore in classification tasks
        ignore_nan: Whether to ignore NaN values in regression tasks
        **meters: Named MetricMeter instances or metric functions

    Examples:
        >>> from danling.metric.functional import accuracy, auroc, auprc, base_preprocess
        >>> meters = MetricMeters(acc=accuracy, auroc=auroc, auprc=auprc, preprocess=base_preprocess)
        >>> meters.update([0.1, 0.8, 0.6, 0.2], [0, 1, 0, 0])
        >>> meters.sum.dict()
        {'acc': 3.0, 'auroc': 4.0, 'auprc': 4.0}
        >>> meters.count.dict()
        {'acc': 4, 'auroc': 4, 'auprc': 4}
        >>> meters['auroc'].update([0.2, 0.8], [0, 1])
        >>> meters.sum.dict()
        {'acc': 3.0, 'auroc': 6.0, 'auprc': 4.0}
        >>> meters.count.dict()
        {'acc': 4, 'auroc': 6, 'auprc': 4}
        >>> meters.update([[0.1, 0.7, 0.3, 0.2], [0.8, 0.4]], [[0, 0, 1, 0], [0, 0]])
        >>> meters.sum.dict()
        {'acc': 6.0, 'auroc': 8.4, 'auprc': 5.5}
        >>> meters.count.dict()
        {'acc': 10, 'auroc': 12, 'auprc': 10}
        >>> meters['auroc'].update([0.4, 0.8, 0.6, 0.2], [0, 1, 1, 0])
        >>> meters.avg.dict()
        {'acc': 0.6, 'auroc': 0.775, 'auprc': 0.55}
        >>> meters.update(dict(loss=""))  # doctest: +ELLIPSIS
        Traceback (most recent call last):
        TypeError: ...update() missing 1 required positional argument: 'target'

    Notes:
        - `MetricMeters` manages multiple `MetricMeter` instances with shared preprocessing
        - Each metric is computed independently but uses the same inputs
        - All meters are updated simultaneously when you call `update()`
        - Individual meters can be accessed like dictionary items or attributes

    See Also:
        - [`AverageMeters`][danling.metric.average_meter.AverageMeters]:
            A container for managing multiple average meters in one object.
        - [`Metrics`][danling.metric.metrics.Metrics]:
            Metric tracker that stores the complete prediction and target history.
    """

    preprocess = None
    ignore_index = -100
    ignore_nan = True

    def __init__(
        self,
        *args,
        preprocess: Callable | None = None,
        ignore_index: int | None = None,
        ignore_nan: bool | None = None,
        **meters,
    ) -> None:
        if args:
            from .metrics import Metrics

            if len(args) == 1 and isinstance(args[0], Metrics):
                metrics = args[0]
                for name, metric in metrics.metrics.items():
                    meters.setdefault(name, metric)
                if preprocess is None:
                    preprocess = metrics.preprocess
                if ignore_index is None:
                    ignore_index = metrics.ignore_index
                if ignore_nan is None:
                    ignore_nan = metrics.ignore_nan
            else:
                for metric in args:
                    if not callable(metric):
                        raise ValueError(f"Expected metric to be callable, but got {type(metric)}")
                    meters.setdefault(metric.__name__, metric)
        self.setattr("preprocess", preprocess)
        if ignore_index is not None:
            self.setattr("ignore_index", ignore_index)
        if ignore_nan is not None:
            self.setattr("ignore_nan", ignore_nan)
        for name, meter in meters.items():
            if callable(meter):
                meters[name] = meter = self._preprocess_callable(meter)
            if not isinstance(meter, MetricMeter):
                raise ValueError(f"Expected {name} to be an instance of MetricMeter, but got {type(meter)}")
        super().__init__(default_factory=None, **meters)  # type: ignore[arg-type]

    def update(  # type: ignore[override] # pylint: disable=W0221
        self,
        input: Tensor | NestedTensor | Sequence,  # pylint: disable=W0622
        target: Tensor | NestedTensor | Sequence,
    ) -> None:
        r"""
        Updates the average and current value in all meters.

        Args:
            input: Input values to compute the metrics.
            target: Target values to compute the metrics.
        """

        preprocessed = False
        if self.preprocess is not None:
            input, target = self.preprocess(input, target, ignore_index=self.ignore_index, ignore_nan=self.ignore_nan)
            preprocessed = True
        for meter in self.values():
            meter.update(input, target, preprocessed=preprocessed)

    def set(self, name: str, meter: MetricMeter | Callable) -> None:  # type: ignore[override] # pylint: disable=W0237
        if callable(meter):
            meter = self._preprocess_callable(meter)
        if not isinstance(meter, MetricMeter):
            raise ValueError(f"Expected meter to be an instance of MetricMeter, but got {type(meter)}")
        super().set(name, meter)

    def _preprocess_callable(self, func: Callable) -> MetricMeter:
        if not callable(func):
            raise ValueError(f"Expected func to be callable, but got {type(func)}")
        if "preprocess" in signature(func).parameters:
            func = partial(func, preprocess=self.preprocess is None)
        return MetricMeter(func, preprocess=self.preprocess, ignore_index=self.ignore_index, ignore_nan=self.ignore_nan)

    def __repr__(self):
        keys = tuple(i for i in self.keys())
        return f"{self.__class__.__name__}{keys}"

update

Python
update(
    input: Tensor | NestedTensor | Sequence,
    target: Tensor | NestedTensor | Sequence,
) -> None

Updates the average and current value in all meters.

Parameters:

Name Type Description Default
input
Tensor | NestedTensor | Sequence

Input values to compute the metrics.

 required
target
Tensor | NestedTensor | Sequence

Target values to compute the metrics.

 required
Source code in danling/metric/metric_meter.py
Python
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
def update(  # type: ignore[override] # pylint: disable=W0221
    self,
    input: Tensor | NestedTensor | Sequence,  # pylint: disable=W0622
    target: Tensor | NestedTensor | Sequence,
) -> None:
    r"""
    Updates the average and current value in all meters.

    Args:
        input: Input values to compute the metrics.
        target: Target values to compute the metrics.
    """

    preprocessed = False
    if self.preprocess is not None:
        input, target = self.preprocess(input, target, ignore_index=self.ignore_index, ignore_nan=self.ignore_nan)
        preprocessed = True
    for meter in self.values():
        meter.update(input, target, preprocessed=preprocessed)

MultiTaskAverageMeters

Bases: MultiTaskDict

Manages multiple average meters in one object for multi-task learning.

Examples:

Python Console Session
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
>>> meters = MultiTaskAverageMeters()
>>> meters.update({"loss": 0.6, "dataset1.cls.auroc": 0.7, "dataset1.reg.r2": 0.8, "dataset2.r2": 0.9})
>>> f"{meters:.4f}"
'loss: 0.6000 (0.6000)\ndataset1.cls.auroc: 0.7000 (0.7000)\ndataset1.reg.r2: 0.8000 (0.8000)\ndataset2.r2: 0.9000 (0.9000)'
>>> meters['loss'].update(0.9, n=1)
>>> f"{meters:.4f}"
'loss: 0.9000 (0.7500)\ndataset1.cls.auroc: 0.7000 (0.7000)\ndataset1.reg.r2: 0.8000 (0.8000)\ndataset2.r2: 0.9000 (0.9000)'
>>> meters.sum.dict()
{'loss': 1.5, 'dataset1': {'cls': {'auroc': 0.7}, 'reg': {'r2': 0.8}}, 'dataset2': {'r2': 0.9}}
>>> meters.count.dict()
{'loss': 2, 'dataset1': {'cls': {'auroc': 1}, 'reg': {'r2': 1}}, 'dataset2': {'r2': 1}}
>>> meters.reset()
MultiTaskAverageMeters(...)
>>> f"{meters:.4f}"
'loss: 0.0000 (nan)\ndataset1.cls.auroc: 0.0000 (nan)\ndataset1.reg.r2: 0.0000 (nan)\ndataset2.r2: 0.0000 (nan)'
>>> meters = MultiTaskAverageMeters(return_average=True)
>>> meters.update({"loss": 0.6, "dataset1.a.auroc": 0.7, "dataset1.b.auroc": 0.8, "dataset2.auroc": 0.9})
>>> f"{meters:.4f}"
'loss: 0.6000 (0.6000)\ndataset1.a.auroc: 0.7000 (0.7000)\ndataset1.b.auroc: 0.8000 (0.8000)\ndataset2.auroc: 0.9000 (0.9000)'
>>> meters.update({"loss": 0.9, "dataset1.a.auroc": 0.8, "dataset1.b.auroc": 0.9, "dataset2.auroc": 1.0})
>>> f"{meters:.4f}"
'loss: 0.9000 (0.7500)\ndataset1.a.auroc: 0.8000 (0.7500)\ndataset1.b.auroc: 0.9000 (0.8500)\ndataset2.auroc: 1.0000 (0.9500)'
See Also
  • MultiTaskMetricMeters: Memory-efficient metric tracker that averages multiple metrics batch-by-batch for multi-task learning.
Source code in danling/metric/average_meter.py
Python
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
class MultiTaskAverageMeters(MultiTaskDict):
    r"""
    Manages multiple average meters in one object for multi-task learning.

    Examples:
        >>> meters = MultiTaskAverageMeters()
        >>> meters.update({"loss": 0.6, "dataset1.cls.auroc": 0.7, "dataset1.reg.r2": 0.8, "dataset2.r2": 0.9})
        >>> f"{meters:.4f}"
        'loss: 0.6000 (0.6000)\ndataset1.cls.auroc: 0.7000 (0.7000)\ndataset1.reg.r2: 0.8000 (0.8000)\ndataset2.r2: 0.9000 (0.9000)'
        >>> meters['loss'].update(0.9, n=1)
        >>> f"{meters:.4f}"
        'loss: 0.9000 (0.7500)\ndataset1.cls.auroc: 0.7000 (0.7000)\ndataset1.reg.r2: 0.8000 (0.8000)\ndataset2.r2: 0.9000 (0.9000)'
        >>> meters.sum.dict()
        {'loss': 1.5, 'dataset1': {'cls': {'auroc': 0.7}, 'reg': {'r2': 0.8}}, 'dataset2': {'r2': 0.9}}
        >>> meters.count.dict()
        {'loss': 2, 'dataset1': {'cls': {'auroc': 1}, 'reg': {'r2': 1}}, 'dataset2': {'r2': 1}}
        >>> meters.reset()  # doctest:+ELLIPSIS
        MultiTaskAverageMeters(...)
        >>> f"{meters:.4f}"
        'loss: 0.0000 (nan)\ndataset1.cls.auroc: 0.0000 (nan)\ndataset1.reg.r2: 0.0000 (nan)\ndataset2.r2: 0.0000 (nan)'
        >>> meters = MultiTaskAverageMeters(return_average=True)
        >>> meters.update({"loss": 0.6, "dataset1.a.auroc": 0.7, "dataset1.b.auroc": 0.8, "dataset2.auroc": 0.9})
        >>> f"{meters:.4f}"
        'loss: 0.6000 (0.6000)\ndataset1.a.auroc: 0.7000 (0.7000)\ndataset1.b.auroc: 0.8000 (0.8000)\ndataset2.auroc: 0.9000 (0.9000)'
        >>> meters.update({"loss": 0.9, "dataset1.a.auroc": 0.8, "dataset1.b.auroc": 0.9, "dataset2.auroc": 1.0})
        >>> f"{meters:.4f}"
        'loss: 0.9000 (0.7500)\ndataset1.a.auroc: 0.8000 (0.7500)\ndataset1.b.auroc: 0.9000 (0.8500)\ndataset2.auroc: 1.0000 (0.9500)'

    See Also:
        - [`MultiTaskMetricMeters`][danling.metric.metric_meter.MultiTaskMetricMeters]:
            Memory-efficient metric tracker that averages multiple metrics batch-by-batch for multi-task learning.
    """  # noqa: E501

    @property
    def sum(self) -> NestedDict[str, float]:
        return NestedDict({key: meter.sum for key, meter in self.all_items()})

    @property
    def count(self) -> NestedDict[str, int]:
        return NestedDict({key: meter.count for key, meter in self.all_items()})

    def update(self, *args: Dict, **values: float) -> None:  # pylint: disable=W0237
        r"""
        Updates the average and current value in all meters.

        Args:
            values: Dict of values to be added to the average.
            n: Number of values to be added.

        Raises:
            ValueError: If the value is not an instance of (int, float, Mapping).
        """  # noqa: E501

        if args:
            if len(args) > 1:
                raise ValueError("Expected only one positional argument, but got multiple.")
            values = args[0].update(values) or args[0] if values else args[0]

        for meter, value in values.items():
            if not isinstance(value, (int, float, Mapping)):
                raise ValueError(f"Expected values to be int, float, or a Mapping, but got {type(value)}")
            self[meter].update(value)

    # evil hack, as the default_factory must not be set to make `NestedDict` happy
    # this have some side effects, it will break attribute style intermediate nested dict auto creation
    # but everything has a price
    def get(self, name: Any, default=None) -> Any:
        if not name.startswith("_") and not name.endswith("_"):
            return self.setdefault(name, AverageMeter())
        return super().get(name, default)

    def set(self, name: str, meter: AverageMeter | AverageMeters) -> None:  # pylint: disable=W0237
        if not isinstance(meter, (AverageMeter, AverageMeters)):
            raise ValueError(
                f"Expected meter to be an instance of AverageMeter or AverageMeters, but got {type(meter)}"
            )
        super().set(name, meter, convert_mapping=False)

update

Python
update(*args: Dict, **values: float) -> None

Updates the average and current value in all meters.

Parameters:

Name Type Description Default
values
float

Dict of values to be added to the average.

 {}
n

Number of values to be added.

 required

Raises:

Type Description
ValueError

If the value is not an instance of (int, float, Mapping).
Source code in danling/metric/average_meter.py
Python
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
def update(self, *args: Dict, **values: float) -> None:  # pylint: disable=W0237
    r"""
    Updates the average and current value in all meters.

    Args:
        values: Dict of values to be added to the average.
        n: Number of values to be added.

    Raises:
        ValueError: If the value is not an instance of (int, float, Mapping).
    """  # noqa: E501

    if args:
        if len(args) > 1:
            raise ValueError("Expected only one positional argument, but got multiple.")
        values = args[0].update(values) or args[0] if values else args[0]

    for meter, value in values.items():
        if not isinstance(value, (int, float, Mapping)):
            raise ValueError(f"Expected values to be int, float, or a Mapping, but got {type(value)}")
        self[meter].update(value)

MultiTaskMetricMeters

Bases: MultiTaskAverageMeters

Examples:

Python Console Session
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
>>> from danling.metric.functional import accuracy
>>> metrics = MultiTaskMetricMeters()
>>> metrics.dataset1.cls = MetricMeters(acc=accuracy)
>>> metrics.dataset2 = MetricMeters(acc=accuracy)
>>> metrics
MultiTaskMetricMeters(<class 'danling.metric.metric_meter.MultiTaskMetricMeters'>,
  ('dataset1'): MultiTaskMetricMeters(<class 'danling.metric.metric_meter.MultiTaskMetricMeters'>,
    ('cls'): MetricMeters('acc',)
  )
  ('dataset2'): MetricMeters('acc',)
)
>>> metrics.update({"dataset1.cls": {"input": [0.2, 0.4, 0.5, 0.7], "target": [0, 1, 0, 1]}, "dataset2": ([0.1, 0.4, 0.6, 0.8], [1, 0, 0, 0])})
>>> f"{metrics:.4f}"
'dataset1.cls: acc: 0.5000 (0.5000)\ndataset2: acc: 0.2500 (0.2500)'
>>> metrics.setattr("return_average", True)
>>> metrics.update({"dataset1.cls": [[0.1, 0.4, 0.6, 0.8], [0, 0, 1, 0]], "dataset2": {"input": [0.2, 0.3, 0.5, 0.7], "target": [0, 0, 0, 1]}})
>>> f"{metrics:.4f}"
'dataset1.cls: acc: 0.7500 (0.6250)\ndataset2: acc: 0.7500 (0.5000)'
>>> metrics.update(dict(loss=""))
Traceback (most recent call last):
ValueError: Metric loss not found in ...
Notes
  • MultiTaskMetricMeters manages nested hierarchies of MetricMeters for multiple tasks/datasets
  • Supports hierarchical access using dot notation or dictionary-style access
  • All metrics are updated simultaneously with a single update() call
  • Provides a structured way to track metrics across different tasks or model components
See Also
  • MultiTaskAverageMeters: A container for managing multiple average meters in one object for multi-task learning.
  • MultiTaskMetrics: Metric tracker that stores the complete prediction and target history for multi-task learning.
Source code in danling/metric/metric_meter.py
Python
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
class MultiTaskMetricMeters(MultiTaskAverageMeters):
    r"""
    Examples:
        >>> from danling.metric.functional import accuracy
        >>> metrics = MultiTaskMetricMeters()
        >>> metrics.dataset1.cls = MetricMeters(acc=accuracy)
        >>> metrics.dataset2 = MetricMeters(acc=accuracy)
        >>> metrics
        MultiTaskMetricMeters(<class 'danling.metric.metric_meter.MultiTaskMetricMeters'>,
          ('dataset1'): MultiTaskMetricMeters(<class 'danling.metric.metric_meter.MultiTaskMetricMeters'>,
            ('cls'): MetricMeters('acc',)
          )
          ('dataset2'): MetricMeters('acc',)
        )
        >>> metrics.update({"dataset1.cls": {"input": [0.2, 0.4, 0.5, 0.7], "target": [0, 1, 0, 1]}, "dataset2": ([0.1, 0.4, 0.6, 0.8], [1, 0, 0, 0])})
        >>> f"{metrics:.4f}"
        'dataset1.cls: acc: 0.5000 (0.5000)\ndataset2: acc: 0.2500 (0.2500)'
        >>> metrics.setattr("return_average", True)
        >>> metrics.update({"dataset1.cls": [[0.1, 0.4, 0.6, 0.8], [0, 0, 1, 0]], "dataset2": {"input": [0.2, 0.3, 0.5, 0.7], "target": [0, 0, 0, 1]}})
        >>> f"{metrics:.4f}"
        'dataset1.cls: acc: 0.7500 (0.6250)\ndataset2: acc: 0.7500 (0.5000)'
        >>> metrics.update(dict(loss=""))  # doctest: +ELLIPSIS
        Traceback (most recent call last):
        ValueError: Metric loss not found in ...

    Notes:
        - `MultiTaskMetricMeters` manages nested hierarchies of MetricMeters for multiple tasks/datasets
        - Supports hierarchical access using dot notation or dictionary-style access
        - All metrics are updated simultaneously with a single `update()` call
        - Provides a structured way to track metrics across different tasks or model components

    See Also:
        - [`MultiTaskAverageMeters`][danling.metric.average_meter.MultiTaskAverageMeters]:
            A container for managing multiple average meters in one object for multi-task learning.
        - [`MultiTaskMetrics`][danling.metric.metrics.MultiTaskMetrics]:
            Metric tracker that stores the complete prediction and target history for multi-task learning.
    """  # noqa: E501

    def __init__(self, *args, **kwargs):
        super().__init__(*args, default_factory=MultiTaskMetricMeters, **kwargs)

    def update(  # type: ignore[override] # pylint: disable=W0221
        self,
        values: Mapping[str, Tuple[Tensor | NestedTensor | Sequence, Tensor | NestedTensor | Sequence]],
    ) -> None:
        r"""
        Updates the average and current value in all meters.

        Args:
            input: Input values to compute the metrics.
            target: Target values to compute the metrics.
        """

        for metric, value in values.items():
            if metric not in self:
                raise ValueError(f"Metric {metric} not found in {self}")
            if isinstance(self[metric], MultiTaskMetricMeters):
                for met in self[metric].all_values():
                    if isinstance(value, Mapping):
                        met.update(**value)
                    elif isinstance(value, Sequence):
                        met.update(*value)
                    else:
                        raise ValueError(f"Expected value to be a Mapping or Sequence, but got {type(value)}")
            elif isinstance(self[metric], (MetricMeters, MetricMeter)):
                if isinstance(value, Mapping):
                    self[metric].update(**value)
                elif isinstance(value, Sequence):
                    self[metric].update(*value)
                else:
                    raise ValueError(f"Expected value to be a Mapping or Sequence, but got {type(value)}")
            else:
                raise ValueError(
                    f"Expected {metric} to be an instance of MultiTaskMetricMeters, MetricMeters, "
                    f"or MetricMeter, but got {type(self[metric])}"
                )

    # MultiTaskAverageMeters.get is hacked
    def get(self, name: Any, default=None) -> Any:
        return MultiTaskDict.get(self, name, default)

    def set(  # pylint: disable=W0237
        self,
        name: str,
        meter: MetricMeter | MetricMeters | Callable,  # type: ignore[override]
    ) -> None:
        from .metrics import Metrics

        if isinstance(meter, Metrics):
            meter = MetricMeters(
                preprocess=meter.preprocess,
                ignore_index=meter.ignore_index,
                ignore_nan=meter.ignore_nan,
                **meter.metrics,
            )
        elif callable(meter):
            meter = MetricMeter(meter, preprocess=None)
        elif not isinstance(meter, (MetricMeter, MetricMeters)):
            raise ValueError(
                f"Expected {meter} to be an instance of MetricMeter or MetricMeters, but got {type(meter)}"
            )
        super().set(name, meter)

update

Python
update(
    values: Mapping[
        str,
        Tuple[
            Tensor | NestedTensor | Sequence,
            Tensor | NestedTensor | Sequence,
        ],
    ]
) -> None

Updates the average and current value in all meters.

Parameters:

Name Type Description Default
input

Input values to compute the metrics.

 required
target

Target values to compute the metrics.

 required
Source code in danling/metric/metric_meter.py
Python
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
def update(  # type: ignore[override] # pylint: disable=W0221
    self,
    values: Mapping[str, Tuple[Tensor | NestedTensor | Sequence, Tensor | NestedTensor | Sequence]],
) -> None:
    r"""
    Updates the average and current value in all meters.

    Args:
        input: Input values to compute the metrics.
        target: Target values to compute the metrics.
    """

    for metric, value in values.items():
        if metric not in self:
            raise ValueError(f"Metric {metric} not found in {self}")
        if isinstance(self[metric], MultiTaskMetricMeters):
            for met in self[metric].all_values():
                if isinstance(value, Mapping):
                    met.update(**value)
                elif isinstance(value, Sequence):
                    met.update(*value)
                else:
                    raise ValueError(f"Expected value to be a Mapping or Sequence, but got {type(value)}")
        elif isinstance(self[metric], (MetricMeters, MetricMeter)):
            if isinstance(value, Mapping):
                self[metric].update(**value)
            elif isinstance(value, Sequence):
                self[metric].update(*value)
            else:
                raise ValueError(f"Expected value to be a Mapping or Sequence, but got {type(value)}")
        else:
            raise ValueError(
                f"Expected {metric} to be an instance of MultiTaskMetricMeters, MetricMeters, "
                f"or MetricMeter, but got {type(self[metric])}"
            )

LRScheduler

Bases: _LRScheduler

General learning rate scheduler.

PyTorch LRScheduler is hard to extend. This class is a wrapper of PyTorch LRScheduler, which provides a more general interface. You only needs to add a new method which calculates a learning rate ratio (range from 0 to 1) with total progress (range from 0 to 1), and everything else will be done automatically.

Moreover, this class has warmup and cooldown built-in. By default, the first 5% and last 20% of training steps will be warmup and cooldown respectively. You can alternate by passing warmup_steps and cooldown_steps, or disable them by setting them to 0.

Parameters:

Name Type Description Default

optimizer

Optimizer

Wrapped optimizer.

 required

total_steps

int

Total number of trainable steps.

 required

final_lr_ratio

Optional[float]

Final learning rate ratio to initial learning rate. Defaults to 1e-3.

 None

final_lr

Optional[float]

Final learning rate.

 None

min_lr

float

Minimal learning rate. Defaults to 1e-9.

 1e-09

strategy

str

Scaling strategy. Defaults to “cosine”.

 'cosine'

warmup_steps

Optional[int]

Number of warmup steps. Defaults to steps // 20.

 None

cooldown_steps

Optional[int]

Number of cooldown steps. Defaults to steps // 5.

 None

last_epoch

int

The index of last epoch. Defaults to -1.

 -1

method

Optional[str]

Method to calculate learning rate given ratio, should be one of “percentile” or “numerical”. Defaults to “percentile” if final_lr_ratio is set, otherwise “numerical”.

 None

Examples:

Python Console Session
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
>>> from danling.optim import LRScheduler
>>> import torch
>>> from torch import optim
>>> optimizer = optim.SGD([{'params': torch.tensor([0])}], lr=1, momentum=0.9)
>>> scheduler = LRScheduler(optimizer, total_steps=5, final_lr_ratio=1e-5, strategy='linear')
>>> lrs = []
>>> for epoch in range(5):
...     lrs.append(scheduler.get_lr()[0])
...     scheduler.step()
>>> [round(lr, 10) for lr in lrs]
[0.1, 0.01, 0.001, 0.0001, 1e-09]
>>> scheduler = LRScheduler(optimizer, total_steps=5, final_lr_ratio=1e-5, strategy='cosine')
>>> lrs = []
>>> for epoch in range(5):
...     lrs.append(scheduler.get_lr()[0])
...     scheduler.step()
>>> [round(lr, 10) for lr in lrs]
[0.3330753446, 0.0187302031, 0.000533897, 3.00232e-05, 1e-09]
>>> scheduler = LRScheduler(optimizer, total_steps=5, final_lr_ratio=1e-5, strategy='linear', method='numerical')
>>> lrs = []
>>> for epoch in range(5):
...     lrs.append(scheduler.get_lr()[0])
...     scheduler.step()
>>> [round(lr, 2) for lr in lrs]
[0.8, 0.6, 0.4, 0.2, 0.0]
Source code in danling/optim/lr_scheduler/lr_scheduler.py
Python
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
class LRScheduler(lr_scheduler._LRScheduler):  # pylint: disable=protected-access
    r"""
    General learning rate scheduler.

    PyTorch LRScheduler is hard to extend.
    This class is a wrapper of PyTorch LRScheduler, which provides a more general interface.
    You only needs to add a new method which calculates a learning rate ratio (range from 0 to 1)
    with total progress (range from 0 to 1), and everything else will be done automatically.

    Moreover, this class has warmup and cooldown built-in.
    By default, the first 5% and last 20% of training steps will be warmup and cooldown respectively.
    You can alternate by passing `warmup_steps` and `cooldown_steps`, or disable them by setting them to 0.

    Args:
        optimizer: Wrapped optimizer.
        total_steps: Total number of trainable steps.
        final_lr_ratio: Final learning rate ratio to initial learning rate.
            Defaults to 1e-3.
        final_lr: Final learning rate.
        min_lr: Minimal learning rate.
            Defaults to 1e-9.
        strategy: Scaling strategy.
            Defaults to "cosine".
        warmup_steps: Number of warmup steps.
            Defaults to `steps // 20`.
        cooldown_steps: Number of cooldown steps.
            Defaults to `steps // 5`.
        last_epoch: The index of last epoch.
            Defaults to -1.
        method: Method to calculate learning rate given ratio, should be one of "percentile" or "numerical".
            Defaults to "percentile" if `final_lr_ratio` is set, otherwise "numerical".

    Examples:
        >>> from danling.optim import LRScheduler
        >>> import torch
        >>> from torch import optim
        >>> optimizer = optim.SGD([{'params': torch.tensor([0])}], lr=1, momentum=0.9)
        >>> scheduler = LRScheduler(optimizer, total_steps=5, final_lr_ratio=1e-5, strategy='linear')
        >>> lrs = []
        >>> for epoch in range(5):
        ...     lrs.append(scheduler.get_lr()[0])
        ...     scheduler.step()
        >>> [round(lr, 10) for lr in lrs]
        [0.1, 0.01, 0.001, 0.0001, 1e-09]
        >>> scheduler = LRScheduler(optimizer, total_steps=5, final_lr_ratio=1e-5, strategy='cosine')
        >>> lrs = []
        >>> for epoch in range(5):
        ...     lrs.append(scheduler.get_lr()[0])
        ...     scheduler.step()
        >>> [round(lr, 10) for lr in lrs]
        [0.3330753446, 0.0187302031, 0.000533897, 3.00232e-05, 1e-09]
        >>> scheduler = LRScheduler(optimizer, total_steps=5, final_lr_ratio=1e-5, strategy='linear', method='numerical')
        >>> lrs = []
        >>> for epoch in range(5):
        ...     lrs.append(scheduler.get_lr()[0])
        ...     scheduler.step()
        >>> [round(lr, 2) for lr in lrs]
        [0.8, 0.6, 0.4, 0.2, 0.0]
    """  # noqa: E501

    def __init__(
        self,
        optimizer: Optimizer,
        total_steps: int,
        final_lr_ratio: Optional[float] = None,
        final_lr: Optional[float] = None,
        min_lr: float = 1e-9,
        strategy: str = "cosine",
        warmup_steps: Optional[int] = None,
        cooldown_steps: Optional[int] = None,
        last_epoch: int = -1,
        method: Optional[str] = None,
    ):
        if total_steps <= 0:
            raise ValueError(f"Total steps must be positive, but got {total_steps}")
        if warmup_steps is None:
            warmup_steps = total_steps // 20
        elif warmup_steps > total_steps:
            raise ValueError(f"Warmup steps must be less than total steps, but got {warmup_steps} > {total_steps}")
        elif warmup_steps < 0:
            raise ValueError(f"Warmup steps must be positive, but got {warmup_steps}")
        if cooldown_steps is None:
            cooldown_steps = total_steps // 5
        elif cooldown_steps > total_steps:
            raise ValueError(f"Cooldown steps must be less than total steps, but got {cooldown_steps} > {total_steps}")
        elif cooldown_steps < 0:
            raise ValueError(f"Cooldown steps must be positive, but got {cooldown_steps}")
        if warmup_steps + cooldown_steps > total_steps:
            raise ValueError(
                "Warmup steps + cooldown steps must be less than total steps, "
                f"but got {warmup_steps} + {cooldown_steps} > {total_steps}"
            )
        if final_lr_ratio is not None:
            if final_lr is not None:
                raise ValueError("Only one of `final_lr_ratio` and `final_lr` should be set, but not both")
            if final_lr_ratio < 0:
                raise ValueError(f"`final_lr_ratio` must be positive, but got {final_lr_ratio}")
            if method is None:
                method = "percentile"
        if final_lr is not None and final_lr < 0:
            raise ValueError(f"`final_lr` must be positive, but got {final_lr}")
        if min_lr < 0:
            raise ValueError(f"`min_lr` must be positive, but got {min_lr}")
        self.strategies = {
            k: v for k, v in self.__class__.__dict__.items() if callable(v) and (not k.startswith("_") or k in "get_lr")
        }
        if strategy not in self.strategies:
            raise ValueError(f"Scaling strategy must be one of {self.strategies.keys()}, but got {strategy}")

        if final_lr_ratio is None and final_lr is None:
            final_lr_ratio = 1e-3
            if method is None:
                method = "percentile"
        if final_lr is not None and min_lr > final_lr:
            min_lr = final_lr
        if method is None:
            method = "numerical"

        self.final_lr_ratio = final_lr_ratio
        self.final_lr = final_lr
        self.total_steps = total_steps
        self.min_lr = min_lr
        self.strategy = strategy
        self.method = method
        self.warmup_steps = warmup_steps
        self.cooldown_steps = cooldown_steps
        self.cooldown_steps_begin = self.total_steps - self.cooldown_steps
        super().__init__(optimizer, last_epoch)

    def get_lr(self) -> List[float]:
        step_count = self._step_count
        if step_count > self.total_steps + 1 or step_count < 1:
            warn(
                f"Step count {step_count} is out of range [1, {self.total_steps + 1}]",
                category=RuntimeWarning,
                stacklevel=2,
            )
        return [self._get_lr(lr, step_count) for lr in self.base_lrs]

    def _get_lr(
        self,
        lr: float,
        step_count: Optional[int] = None,
        progress: Optional[float] = None,
        warmup_ratio: Optional[float] = None,
        cooldown_ratio: Optional[float] = None,
        method: Optional[str] = None,
    ) -> float:
        method = method or self.method
        step_count = step_count or self._step_count
        progress = progress or min(max(step_count / self.total_steps, 0.0), 1.0)
        final_lr = self.final_lr if self.final_lr is not None else lr * self.final_lr_ratio  # type: ignore[operator]
        ratio = getattr(self, self.strategy)(progress)
        if method == "percentile":
            lr *= pow(final_lr / lr, ratio)
        elif method == "numerical":
            lr = (1 - ratio) * (lr - final_lr) + final_lr
        else:
            raise ValueError(f"Method must be one of ['percentile', 'numerical'], but got {method}")
        if self.warmup_steps > step_count > 0:
            warmup_ratio = warmup_ratio or step_count / self.warmup_steps
            lr = warmup_ratio * (lr - self.min_lr) + self.min_lr
        elif self.cooldown_steps > 0 and step_count > self.cooldown_steps_begin:
            cooldown_ratio = cooldown_ratio or 1 - (step_count - self.cooldown_steps_begin) / self.cooldown_steps
            lr = cooldown_ratio * (lr - self.min_lr) + self.min_lr
        return max(self.min_lr, lr)

    def linear(self, progress: float) -> float:
        return progress

    def cosine(self, progress: float) -> float:
        return 1 - ((1 + cos(pi * progress)) / 2)

    def constant(self, progress: float) -> float:  # pylint: disable=unused-argument
        return 0.0

    def __repr__(self) -> str:
        return (
            f"{self.__class__.__name__}({self.strategy}, method={self.method}, "
            f"final_lr_ratio={self.final_lr_ratio}, total_steps={self.total_steps}, "
            f"warmup_steps={self.warmup_steps}, cooldown_steps={self.cooldown_steps})"
        )

AccelerateRunner

Bases: TorchRunner, Accelerator

Set up everything for running a job with 🤗 accelerate.

AccelerateRunner extends the Accelerator class to provide a more user-friendly and consistent interface.

AccelerateRunner provides the most easy-to-use interface for distributed training, but it can be slow, and not very flexible. Read the documentation of Accelerator for more details.

Source code in danling/runner/accelerate_runner.py
Python
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
class AccelerateRunner(TorchRunner, Accelerator):  # pylint: disable=too-many-public-methods
    r"""
    Set up everything for running a job with 🤗 [`accelerate`](https://huggingface.co/docs/accelerate/).

    `AccelerateRunner` extends the [`Accelerator`][accelerate.Accelerator] class to provide a more user-friendly
    and consistent interface.

    `AccelerateRunner` provides the most easy-to-use interface for distributed training, but it can be slow, and not
    very flexible.
    Read the documentation of [`Accelerator`][accelerate.Accelerator] for more details.
    """

    _accelerate: FlatDict | None = None

    def __init__(self, config: Config) -> None:
        ac.check()
        TorchRunner.__init__(self, config)
        Accelerator.__init__(self, **self.accelerate)
        if self.distributed:
            object_list = [self.id, self.timestamp]
            dist.broadcast_object_list(object_list)
            self.id, self.timestamp = object_list

    def __post_init__(self) -> None:
        BaseRunner.__post_init__(self)
        self.project_configuration.set_directories(self.dir)
        if self.datasets:
            self.build_dataloaders()
        if self.config.get("log_interval") is None:
            self.config.log_interval = (
                max(ceil(max(len(d) for d in self.dataloaders.values()) / 10), 1) if self.dataloaders else 1
            )
        self.model, self.criterion, self.optimizer, self.scheduler = self.prepare(
            self.model, self.criterion, self.optimizer, self.scheduler
        )

    def train_step(self, data) -> torch.Tensor:
        with self.autocast(), self.accumulate():
            input = data["input"] if isinstance(data, Mapping) else data[0]
            target = data["target"] if isinstance(data, Mapping) else data[1]
            pred = self.model(**input) if isinstance(input, Mapping) else self.model(input)
            loss = self.criterion(pred, target)
            if self.metrics is not None:
                self.metrics.update(pred.squeeze(-1), target)
            self.advance(loss)
        return loss

    def advance(self, loss) -> None:
        r"""
        Backward loss and step optimizer & scheduler.

        Args:
            loss: The loss tensor from which to backpropagate.
        """

        self.backward(loss)
        if self.sync_gradients:
            if self.config.get("max_grad_value") is not None:
                self.clip_grad_value_(self.model.parameters(), self.config["max_grad_value"])
            if self.config.get("max_grad_norm") is not None:
                self.clip_grad_norm_(self.model.parameters(), self.config["max_grad_norm"])
        self.optimizer.step()
        if self.scheduler is not None:
            self.scheduler.step()
        if self.ema is not None:
            self.ema.update()
        self.optimizer.zero_grad()
        self.config.steps = self.step

    def unwrap(self, model: nn.Module) -> nn.Module:
        return self.unwrap_model(model)

    @property
    def accelerate(self) -> FlatDict:
        if self._accelerate is None:
            self._accelerate = self.get_accelerate_config(self.config)
        return self._accelerate

    @accelerate.setter
    def accelerate(self, config: FlatDict) -> None:
        self._accelerate = config

    @property
    def deepspeed(self) -> dict | None:
        if self.state.deepspeed_plugin is not None:
            return self.state.deepspeed_plugin.deepspeed_config
        return None

    @contextmanager
    def accumulate(self, *models: nn.Module):
        if not models:
            models = (self.model,)
        yield Accelerator.accumulate(self, *models)

    @property
    def device(self) -> torch.device:
        return self.state.device

    @property
    def world_size(self) -> int:
        if "state" in self.__dict__:
            return self.state.num_processes
        return 1

    @property
    def rank(self) -> int:
        if "state" in self.__dict__:
            return self.state.process_index
        return 0

    @property
    def local_rank(self) -> int:
        if "state" in self.__dict__:
            return self.state.local_process_index
        return 0

    @cached_property
    def accum_steps(self) -> int:
        return self.gradient_accumulation_steps

    def get_accelerate_config(self, config) -> FlatDict:
        accelerate = FlatDict()
        if "accelerate" in config:
            accelerate.update(config.accelerate)
        if "precision" in config:
            accelerate.mixed_precision = config.precision
        if "dynamo" in config:
            accelerate.dynamo_backend = config.dynamo.upper()
        if "accum_steps" in config:
            accelerate.gradient_accumulation_steps = config.accum_steps
        if "kwargs_handlers" not in accelerate:
            accelerate.kwargs_handlers = []
        # Must NOT set project_dir here as timestamp is not synced yet
        # config.project_dir = self.dir
        if os.getenv("ACCELERATE_USE_DEEPSPEED", "false").lower() == "true":
            deepspeed_config = config.get("deepspeed", os.getenv("ACCELERATE_DEEPSPEED_CONFIG_FILE"))
            accelerate.deepspeed_plugin = DeepSpeedPlugin(hf_ds_config=self.get_deepspeed_config(deepspeed_config))
        return accelerate

    def build_dataloaders(self):
        datasets = {k: d for k, d in self.datasets.items() if k not in self.dataloaders}
        default_kwargs = self.config.setdefault("dataloader", NestedDict())
        dataloader_kwargs = NestedDict({k: default_kwargs.pop(k) for k in self.datasets if k in default_kwargs})
        for k, d in datasets.items():
            dataloader_kwargs.setdefault(k, NestedDict())
            dataloader_kwargs[k].merge(default_kwargs, overwrite=False)
            dataloader_kwargs[k].setdefault("shuffle", getattr(d, "train", True))
            dataloader_kwargs[k].setdefault("drop_last", not getattr(d, "train", True))
            self.dataloaders[k] = utils.data.DataLoader(d, collate_fn=self.collate_fn, **dataloader_kwargs[k])
        default_kwargs.update(dataloader_kwargs)
        for k, d in self.dataloaders.items():
            self.dataloaders[k] = self.prepare(d)

advance

Python
advance(loss) -> None

Backward loss and step optimizer & scheduler.

Parameters:

Name Type Description Default
loss

The loss tensor from which to backpropagate.

 required
Source code in danling/runner/accelerate_runner.py
Python
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
def advance(self, loss) -> None:
    r"""
    Backward loss and step optimizer & scheduler.

    Args:
        loss: The loss tensor from which to backpropagate.
    """

    self.backward(loss)
    if self.sync_gradients:
        if self.config.get("max_grad_value") is not None:
            self.clip_grad_value_(self.model.parameters(), self.config["max_grad_value"])
        if self.config.get("max_grad_norm") is not None:
            self.clip_grad_norm_(self.model.parameters(), self.config["max_grad_norm"])
    self.optimizer.step()
    if self.scheduler is not None:
        self.scheduler.step()
    if self.ema is not None:
        self.ema.update()
    self.optimizer.zero_grad()
    self.config.steps = self.step

BaseRunner

Base class for DanLing runners.

The BaseRunner class provides a complete infrastructure for managing:

  1. Experiment Environment: Sets up reproducible training environment with:

    • Random seed management
    • Deterministic execution options
    • Configurable logging systems

  2. Experiment Lifecycle: Handles the complete experiment workflow:

    • Model, optimizer, and scheduler setup
    • Dataset and dataloader management
    • Training, evaluation, and inference loops
    • Checkpointing and restoration

  3. Experiment Tracking: Provides comprehensive experiment monitoring:

    • Progress tracking
    • Metrics computation and logging
    • Results collection and comparison
    • Best model selection

  4. Distributed Training: Supports various parallel training configurations:

    • Multi-process distributed training
    • Gradient accumulation
    • Mixed precision training

This class defines standard attributes and methods that can be extended by more specialized runners. It is designed to be modular and configurable to support various experiment requirements.

ID:

Name Type Description
timestamp str

Time string of when the runner was created. Format: WDHMSUUU- (Week, Weekday, Hour, Minute, Second, Microsecond)
name str

Human-readable name combining experiment and run names ({experiment_name}-{run_name}).
id str

Unique identifier generated from experiment and run IDs.
uuid UUID

UUID5 generated from the run ID and combined ID.
See Also

Config: Configuration for runners.

RunnerMode: Enumeration of runner modes.

Model:

Name Type Description
model Callable | None

The model being trained/evaluated.
criterion Callable | None

Loss function used for optimization.
optimizer Any | None

Optimizer used for parameter updates.
scheduler Any | None

Learning rate scheduler for adaptive learning rates.

Data:

Name Type Description
datasets FlatDict

Dictionary mapping split names to dataset objects.
datasamplers FlatDict

Dictionary mapping split names to data samplers.
dataloaders FlatDict

Dictionary mapping split names to dataloader objects.
split str | None

Current active data split.
batch_size int

Number of samples per batch for current split.
batch_size_equivalent int

Total effective batch size accounting for distribution.

Progress:

Name Type Description
progress float

Running progress as a value between 0 and 1.
steps int

Current step count.
epochs int

Current epoch count.
total_steps int

Total steps to be performed.
total_epochs int

Total epochs to be performed.

Results:

Name Type Description
results NestedDict

Hierarchical results organized by epoch, split, and metric.
latest_result NestedDict | None

Most recent evaluation results.
best_result NestedDict | None

Best evaluation results achieved.
scores FlatDict | None

Main metric values used for model comparison.
latest_score float | None

Most recent score.
best_score float | None

Best score achieved.
is_best bool

Whether the latest result is the best so far.

I/O:

Name Type Description
dir str

Root directory for all experiment outputs.
checkpoint_dir str

Directory for saving model checkpoints.
log_file str

Path to the log file.

Distributed Training:

Name Type Description
world_size int

Number of processes in distributed training.
rank int

Global process rank.
local_rank int

Local process rank on current node.
distributed bool

Whether running in distributed mode.
is_main_process bool

Whether current process is the main process.

Logging and Visualization:

Name Type Description
meters AverageMeters

Meters for tracking running averages of metrics.
metrics Metrics | MetricMeters | None

Metric computation objects.
logger Logger | None

Logger for capturing runtime information.
writer Any | None

Writer for visualizing metrics (e.g., TensorBoard).
Source code in danling/runner/base_runner.py
Python
  57
  58
  59
  60
  61
  62
  63
  64
  65
  66
  67
  68
  69
  70
  71
  72
  73
  74
  75
  76
  77
  78
  79
  80
  81
  82
  83
  84
  85
  86
  87
  88
  89
  90
  91
  92
  93
  94
  95
  96
  97
  98
  99
 100
 101
 102
 103
 104
 105
 106
 107
 108
 109
 110
 111
 112
 113
 114
 115
 116
 117
 118
 119
 120
 121
 122
 123
 124
 125
 126
 127
 128
 129
 130
 131
 132
 133
 134
 135
 136
 137
 138
 139
 140
 141
 142
 143
 144
 145
 146
 147
 148
 149
 150
 151
 152
 153
 154
 155
 156
 157
 158
 159
 160
 161
 162
 163
 164
 165
 166
 167
 168
 169
 170
 171
 172
 173
 174
 175
 176
 177
 178
 179
 180
 181
 182
 183
 184
 185
 186
 187
 188
 189
 190
 191
 192
 193
 194
 195
 196
 197
 198
 199
 200
 201
 202
 203
 204
 205
 206
 207
 208
 209
 210
 211
 212
 213
 214
 215
 216
 217
 218
 219
 220
 221
 222
 223
 224
 225
 226
 227
 228
 229
 230
 231
 232
 233
 234
 235
 236
 237
 238
 239
 240
 241
 242
 243
 244
 245
 246
 247
 248
 249
 250
 251
 252
 253
 254
 255
 256
 257
 258
 259
 260
 261
 262
 263
 264
 265
 266
 267
 268
 269
 270
 271
 272
 273
 274
 275
 276
 277
 278
 279
 280
 281
 282
 283
 284
 285
 286
 287
 288
 289
 290
 291
 292
 293
 294
 295
 296
 297
 298
 299
 300
 301
 302
 303
 304
 305
 306
 307
 308
 309
 310
 311
 312
 313
 314
 315
 316
 317
 318
 319
 320
 321
 322
 323
 324
 325
 326
 327
 328
 329
 330
 331
 332
 333
 334
 335
 336
 337
 338
 339
 340
 341
 342
 343
 344
 345
 346
 347
 348
 349
 350
 351
 352
 353
 354
 355
 356
 357
 358
 359
 360
 361
 362
 363
 364
 365
 366
 367
 368
 369
 370
 371
 372
 373
 374
 375
 376
 377
 378
 379
 380
 381
 382
 383
 384
 385
 386
 387
 388
 389
 390
 391
 392
 393
 394
 395
 396
 397
 398
 399
 400
 401
 402
 403
 404
 405
 406
 407
 408
 409
 410
 411
 412
 413
 414
 415
 416
 417
 418
 419
 420
 421
 422
 423
 424
 425
 426
 427
 428
 429
 430
 431
 432
 433
 434
 435
 436
 437
 438
 439
 440
 441
 442
 443
 444
 445
 446
 447
 448
 449
 450
 451
 452
 453
 454
 455
 456
 457
 458
 459
 460
 461
 462
 463
 464
 465
 466
 467
 468
 469
 470
 471
 472
 473
 474
 475
 476
 477
 478
 479
 480
 481
 482
 483
 484
 485
 486
 487
 488
 489
 490
 491
 492
 493
 494
 495
 496
 497
 498
 499
 500
 501
 502
 503
 504
 505
 506
 507
 508
 509
 510
 511
 512
 513
 514
 515
 516
 517
 518
 519
 520
 521
 522
 523
 524
 525
 526
 527
 528
 529
 530
 531
 532
 533
 534
 535
 536
 537
 538
 539
 540
 541
 542
 543
 544
 545
 546
 547
 548
 549
 550
 551
 552
 553
 554
 555
 556
 557
 558
 559
 560
 561
 562
 563
 564
 565
 566
 567
 568
 569
 570
 571
 572
 573
 574
 575
 576
 577
 578
 579
 580
 581
 582
 583
 584
 585
 586
 587
 588
 589
 590
 591
 592
 593
 594
 595
 596
 597
 598
 599
 600
 601
 602
 603
 604
 605
 606
 607
 608
 609
 610
 611
 612
 613
 614
 615
 616
 617
 618
 619
 620
 621
 622
 623
 624
 625
 626
 627
 628
 629
 630
 631
 632
 633
 634
 635
 636
 637
 638
 639
 640
 641
 642
 643
 644
 645
 646
 647
 648
 649
 650
 651
 652
 653
 654
 655
 656
 657
 658
 659
 660
 661
 662
 663
 664
 665
 666
 667
 668
 669
 670
 671
 672
 673
 674
 675
 676
 677
 678
 679
 680
 681
 682
 683
 684
 685
 686
 687
 688
 689
 690
 691
 692
 693
 694
 695
 696
 697
 698
 699
 700
 701
 702
 703
 704
 705
 706
 707
 708
 709
 710
 711
 712
 713
 714
 715
 716
 717
 718
 719
 720
 721
 722
 723
 724
 725
 726
 727
 728
 729
 730
 731
 732
 733
 734
 735
 736
 737
 738
 739
 740
 741
 742
 743
 744
 745
 746
 747
 748
 749
 750
 751
 752
 753
 754
 755
 756
 757
 758
 759
 760
 761
 762
 763
 764
 765
 766
 767
 768
 769
 770
 771
 772
 773
 774
 775
 776
 777
 778
 779
 780
 781
 782
 783
 784
 785
 786
 787
 788
 789
 790
 791
 792
 793
 794
 795
 796
 797
 798
 799
 800
 801
 802
 803
 804
 805
 806
 807
 808
 809
 810
 811
 812
 813
 814
 815
 816
 817
 818
 819
 820
 821
 822
 823
 824
 825
 826
 827
 828
 829
 830
 831
 832
 833
 834
 835
 836
 837
 838
 839
 840
 841
 842
 843
 844
 845
 846
 847
 848
 849
 850
 851
 852
 853
 854
 855
 856
 857
 858
 859
 860
 861
 862
 863
 864
 865
 866
 867
 868
 869
 870
 871
 872
 873
 874
 875
 876
 877
 878
 879
 880
 881
 882
 883
 884
 885
 886
 887
 888
 889
 890
 891
 892
 893
 894
 895
 896
 897
 898
 899
 900
 901
 902
 903
 904
 905
 906
 907
 908
 909
 910
 911
 912
 913
 914
 915
 916
 917
 918
 919
 920
 921
 922
 923
 924
 925
 926
 927
 928
 929
 930
 931
 932
 933
 934
 935
 936
 937
 938
 939
 940
 941
 942
 943
 944
 945
 946
 947
 948
 949
 950
 951
 952
 953
 954
 955
 956
 957
 958
 959
 960
 961
 962
 963
 964
 965
 966
 967
 968
 969
 970
 971
 972
 973
 974
 975
 976
 977
 978
 979
 980
 981
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
class BaseRunner(metaclass=RunnerMeta):  # pylint: disable=too-many-public-methods
    r"""
    Base class for DanLing runners.

    The `BaseRunner` class provides a complete infrastructure for managing:

    1. **Experiment Environment**: Sets up reproducible training environment with:
        - Random seed management
        - Deterministic execution options
        - Configurable logging systems

    2. **Experiment Lifecycle**: Handles the complete experiment workflow:
        - Model, optimizer, and scheduler setup
        - Dataset and dataloader management
        - Training, evaluation, and inference loops
        - Checkpointing and restoration

    3. **Experiment Tracking**: Provides comprehensive experiment monitoring:
        - Progress tracking
        - Metrics computation and logging
        - Results collection and comparison
        - Best model selection

    4. **Distributed Training**: Supports various parallel training configurations:
        - Multi-process distributed training
        - Gradient accumulation
        - Mixed precision training

    This class defines standard attributes and methods that can be extended by more
    specialized runners. It is designed to be modular and configurable to support
    various experiment requirements.

    Attributes: ID:
        timestamp: Time string of when the runner was created.
            Format: `WDHMSUUU-` (Week, Weekday, Hour, Minute, Second, Microsecond)
        name: Human-readable name combining experiment and run names (`{experiment_name}-{run_name}`).
        id: Unique identifier generated from experiment and run IDs.
        uuid: UUID5 generated from the run ID and combined ID.

    See Also:
        [`Config`][danling.runner.Config]: Configuration for runners.

        [`RunnerMode`][danling.runner.utils.RunnerMode]: Enumeration of runner modes.

    Attributes: Model:
        model: The model being trained/evaluated.
        criterion: Loss function used for optimization.
        optimizer: Optimizer used for parameter updates.
        scheduler: Learning rate scheduler for adaptive learning rates.

    Attributes: Data:
        datasets: Dictionary mapping split names to dataset objects.
        datasamplers: Dictionary mapping split names to data samplers.
        dataloaders: Dictionary mapping split names to dataloader objects.
        split: Current active data split.
        batch_size: Number of samples per batch for current split.
        batch_size_equivalent: Total effective batch size accounting for distribution.

    Attributes: Progress:
        progress: Running progress as a value between 0 and 1.
        steps: Current step count.
        epochs: Current epoch count.
        total_steps: Total steps to be performed.
        total_epochs: Total epochs to be performed.

    Attributes: Results:
        results: Hierarchical results organized by epoch, split, and metric.
        latest_result: Most recent evaluation results.
        best_result: Best evaluation results achieved.
        scores: Main metric values used for model comparison.
        latest_score: Most recent score.
        best_score: Best score achieved.
        is_best: Whether the latest result is the best so far.

    Attributes: I/O:
        dir: Root directory for all experiment outputs.
        checkpoint_dir: Directory for saving model checkpoints.
        log_file: Path to the log file.

    Attributes: Distributed Training:
        world_size: Number of processes in distributed training.
        rank: Global process rank.
        local_rank: Local process rank on current node.
        distributed: Whether running in distributed mode.
        is_main_process: Whether current process is the main process.

    Attributes: Logging and Visualization:
        meters: Meters for tracking running averages of metrics.
        metrics: Metric computation objects.
        logger: Logger for capturing runtime information.
        writer: Writer for visualizing metrics (e.g., TensorBoard).
    """

    # DO NOT set default value in class, as they won't be stored in `__dict__`.

    timestamp: str

    _mode: RunnerMode
    _config: Config
    inited: bool = False

    model: Callable | None = None
    ema: Callable | None = None
    criterion: Callable | None = None
    optimizer: Any | None = None
    scheduler: Any | None = None

    datasets: FlatDict
    datasamplers: FlatDict
    dataloaders: FlatDict
    split: str | None = None
    _train_splits: list[str] | None = None
    _evaluate_splits: list[str] | None = None

    results: NestedDict
    meters: AverageMeters
    metrics: Metrics | MetricMeters | None = None
    train_metrics: Metrics | MetricMeters | None = None
    evaluate_metrics: Metrics | MetricMeters | None = None
    logger: logging.Logger | None = None
    writer: Any | None = None

    def __init__(self, config: Config) -> None:
        self.timestamp = get_time_str()
        if "datasets" not in self.__dict__:
            self.datasets = FlatDict()
        if "datasamplers" not in self.__dict__:
            self.datasamplers = FlatDict()
        if "dataloaders" not in self.__dict__:
            self.dataloaders = FlatDict()
        if "results" not in self.__dict__:
            self.results = NestedDict()
        self.meters = AverageMeters()
        self._mode = RunnerMode.train
        # must init config at last to avoid name conflicts
        if not isinstance(config, Config):
            config = Config(config)
        self._config = config
        self.init_distributed()
        self.inited = True
        if "checkpoint" in config:
            self.load_config(config["checkpoint"])

    def __post_init__(self):
        if self.config.seed is not None:
            self.set_seed()
        if self.config.deterministic:
            self.set_deterministic()
        if self.config.log:
            self.init_logging()
        self.init_print()
        if self.config.tensorboard:
            self.init_tensorboard()

    def init_distributed(self) -> None:
        r"""
        Initialise distributed running environment.
        """

    @on_main_process
    def init_logging(self) -> None:
        r"""
        Set up logging.
        """

        # Why is setting up proper logging so !@?#! ugly?
        logging.config.dictConfig(
            {
                "version": 1,
                "disable_existing_loggers": False,
                "formatters": {
                    "standard": {"format": "%(asctime)s [%(levelname)s] %(name)s: %(message)s"},
                },
                "handlers": {
                    "stdout": {
                        "level": "INFO",
                        "formatter": "standard",
                        "class": "logging.StreamHandler",
                        "stream": "ext://sys.stdout",
                    },
                    "logfile": {
                        "level": "DEBUG",
                        "formatter": "standard",
                        "class": "logging.FileHandler",
                        "filename": self.log_file,
                        "mode": "a",
                    },
                },
                "loggers": {
                    "": {
                        "handlers": ["stdout", "logfile"],
                        "level": "DEBUG",
                        "propagate": True,
                    },
                },
            }
        )
        logging.captureWarnings(True)
        self.logger = logging.getLogger("runner")
        self.logger.flush = lambda: [h.flush() for h in self.logger.handlers]  # type: ignore

    def init_print(self, process: int = 0) -> None:
        r"""
        Set up `print`.

        Only print on a specific `process` or when `force = True`.

        Args:
            process: The process to `print` on.

        Note:
            If `self.config.log = True`, the default `print` function will be override by `logging.info`.
        """

        logger = logging.getLogger("print")
        logger.flush = lambda: [h.flush for h in logger.handlers]  # type: ignore
        import builtins as __builtin__  # pylint: disable=C0415

        builtin_print = __builtin__.print

        @catch
        def print(*args, force=False, end="\n", file=None, flush=False, **kwargs):  # pylint: disable=redefined-builtin
            if self.rank == process or force:
                if self.config.log:
                    if not args:
                        args = [""]
                    logger.info(*args, **kwargs)
                else:
                    builtin_print(*args, end=end, file=file, flush=flush, **kwargs)

        __builtin__.print = print

    @on_main_process
    def init_tensorboard(self, *args, **kwargs) -> None:
        r"""
        Set up Tensoraoard SummaryWriter.
        """
        raise NotImplementedError

    def set_seed(self, seed: int = None, bias: int = None) -> int:  # type: ignore[assignment]
        r"""
        Set random seeds for reproducibility across the application.

        This method configures the random number generators for various libraries
        to ensure reproducible behavior in experiments. In the base implementation,
        it sets seeds for Python's built-in `random` module and NumPy's random
        generators if available.

        Derived runners typically extend this method to set seeds for additional
        libraries (e.g., PyTorch, TensorFlow).

        Args:
            seed: The base random seed to use.
                If None, uses the value from `self.config.seed`.

            bias: An offset to add to the seed for process-specific randomness.
                - If specified, adds this value to the seed
                - If None, uses `self.rank` (recommended for distributed training)
                - If False, disables bias (all processes use identical seed)

                Using a per-process bias is important in distributed training to avoid
                all processes applying identical data augmentation.

        Returns:
            int: The actual seed that was set (after applying bias)

        Examples:
            # Set seed from config
            runner.set_seed()

            # Set specific seed
            runner.set_seed(42)

            # Set seed with explicit bias
            runner.set_seed(42, bias=10)

            # Set same seed for all processes (not recommended for data augmentation)
            runner.set_seed(42, bias=False)

        Note:
            - Setting seeds is essential for reproducible experiments
            - Different biases across processes ensure diverse data augmentation
            - Subclasses should call `super().set_seed()` if overriding
        """

        seed = seed or self.config.seed  # type: ignore[assignment]
        bias = bias or self.rank
        if bias:
            seed += bias
        if np_random is not None:
            np_random.seed(seed)
        random.seed(seed)
        return seed

    def set_deterministic(self) -> None:
        r"""
        Set up deterministic.
        """

        raise NotImplementedError

    def scale_lr(
        self,
        lr: float,
        lr_scale_factor: float | None = None,
        batch_size_base: int | None = None,
    ) -> float:
        r"""
        Scale learning rate according to [linear scaling rule](https://arxiv.org/abs/1706.02677).
        """

        if lr_scale_factor in self.config:
            lr_scale_factor = self.config.lr_scale_factor

        if lr_scale_factor is None:
            if batch_size_base is None:
                batch_size_base = getattr(self, "batch_size_base", None)
                if batch_size_base is None:
                    raise ValueError("batch_size_base must be specified to auto scale lr")
            lr_scale_factor = self.batch_size_equivalent / batch_size_base
        elif batch_size_base is not None:
            warn(
                "batch_size_base will be ignored if lr_scale_factor is specified", category=RuntimeWarning, stacklevel=2
            )
        lr = lr * lr_scale_factor
        self.config.lr_scale_factor = lr_scale_factor
        return lr

    def advance(self, loss, *args, **kwargs) -> None:
        r"""
        Backward loss and step optimizer & scheduler.

        Args:
            loss: loss.
        """

        raise NotImplementedError

    def state_dict(self, cls: Callable = dict) -> Mapping:
        r"""
        Return dict of all attributes for checkpoint.
        """

        return cls(self.config)

    def dict(self, cls: Callable = dict) -> Mapping:
        r"""
        Convert config to Mapping.

        Args:
            cls: Target `clc to convert to.
        """

        return self.config.dict(cls)

    @catch
    def save(self, obj: Any, file: PathStr, main_process_only: bool = True, *args, **kwargs) -> File:
        r"""
        Save any file with supported extensions.

        `Runner.save` internally calls `dl.save`,
        but with additional arguments to allow it save only on the main process.
        Moreover, any error raised by `Runner.save` will be caught and logged.
        """

        if main_process_only and self.is_main_process or not main_process_only:
            return save(obj, file, *args, **kwargs)
        return file

    @staticmethod
    def load(file: PathStr, *args, **kwargs) -> Any:
        r"""
        Load any file with supported extensions.

        `Runner.load` is identical to `dl.load`.
        """

        return load(file, *args, **kwargs)

    @catch
    def json(self, file: File, main_process_only: bool = True, *args, **kwargs) -> None:  # pylint: disable=R1710
        r"""
        Dump Runner config to json file.
        """

        if main_process_only and self.is_main_process or not main_process_only:
            return self.config.json(file, *args, **kwargs)

    @classmethod
    def from_json(cls, file: File, *args, **kwargs) -> BaseRunner:
        r"""
        Construct Runner from json file.

        This function calls `self.from_jsons()` to construct object from json string.
        You may overwrite `from_jsons` in case something is not json serializable.
        """

        with FlatDict.open(file) as fp:
            return cls.from_jsons(fp.read(), *args, **kwargs)

    def jsons(self, *args, **kwargs) -> str:
        r"""
        Dump Runner config to json string.
        """

        return self.config.jsons(*args, **kwargs)

    @classmethod
    def from_jsons(cls, string: str, *args, **kwargs) -> BaseRunner:
        r"""
        Construct Runner from json string.
        """

        return cls(Config.from_jsons(string, *args, **kwargs))

    @catch
    def yaml(self, file: File, main_process_only: bool = True, *args, **kwargs) -> None:  # pylint: disable=R1710
        r"""
        Dump Runner config to yaml file.
        """

        if main_process_only and self.is_main_process or not main_process_only:
            return self.config.yaml(file, *args, **kwargs)

    @classmethod
    def from_yaml(cls, file: File, *args, **kwargs) -> BaseRunner:
        r"""
        Construct Runner from yaml file.

        This function calls `self.from_yamls()` to construct object from yaml string.
        You may overwrite `from_yamls` in case something is not yaml serializable.
        """

        with FlatDict.open(file) as fp:
            return cls.from_yamls(fp.read(), *args, **kwargs)

    def yamls(self, *args, **kwargs) -> str:
        r"""
        Dump Runner config to yaml string.
        """

        return self.config.yamls(*args, **kwargs)

    @classmethod
    def from_yamls(cls, string: str, *args, **kwargs) -> BaseRunner:
        r"""
        Construct Runner from yaml string.
        """

        return cls(Config.from_yamls(string, *args, **kwargs))

    def check_dir(self, action: str = "warn") -> bool:
        r"""
        Check if `self.dir` is not empty.

        Args:
            action (str): The action to perform if `self.dir` is not empty.
            Can be one of ("warn", "raise", "ignore"), default is "warn".
        """

        if action and action not in ("warn", "raise", "ignore"):
            raise ValueError(f"action should be one of warn, raise or ignore, but got {action}")
        if os.listdir(self.dir):
            if action == "warn":
                warn(
                    f"Directory `{self.dir}` is not empty",
                    category=RuntimeWarning,
                    stacklevel=2,
                )
            if action == "raise":
                raise RuntimeError(f"Directory `{self.dir}` is not empty")
            return False
        return True

    @catch
    @on_main_process
    def save_checkpoint(self, name: str = "latest", epochs: int | None = None, save_best: bool = True) -> None:
        r"""
        Save the current model state and runner configuration to a checkpoint file.

        This method saves the complete state of the runner, including model weights,
        optimizer state, scheduler state, and configuration to the checkpoint directory.
        The method handles various checkpoint naming strategies based on parameters.

        Args:
            name: Base name of the checkpoint file (without extension).
                Defaults to `"latest"`, which creates a checkpoint at `{checkpoint_dir}/latest.pth`.
            epochs: Epoch number to record in the checkpoint.
                If None, uses `self.epochs` (current epoch counter).
            save_best: Whether to also save a copy of the checkpoint as `best.pth`
                when the current model performance is the best so far (`self.is_best` is True).
                Defaults to True.

        Note:
            - Checkpoints are saved only on the main process in distributed training.
            - The format is determined by the extension, defaulting to `.pth`.
            - The checkpoint includes the complete runner state from `self.state_dict()`.
            - Periodic checkpoints are saved based on `save_interval` configuration:
              If `self.config.save_interval > 0` and `(epochs + 1) % save_interval == 0`,
              the checkpoint will be copied to `{checkpoint_dir}/epoch-{epochs}.pth`.

        See Also:
            [`load_checkpoint`][danling.BaseRunner.load_checkpoint]: Load a saved checkpoint.
            [`state_dict`][danling.BaseRunner.state_dict]: Get the state dictionary to save.
        """

        epochs = epochs or self.epochs
        save_interval = self.config.get("save_interval", -1)
        latest_path = os.path.join(self.checkpoint_dir, f"{name}.pth")
        self.save(self.state_dict(), latest_path)
        if save_interval > 0 and (epochs + 1) % save_interval == 0:
            save_path = os.path.join(self.checkpoint_dir, f"epoch-{epochs}.pth")
            shutil.copy(latest_path, save_path)
        if save_best and self.is_best:
            best_path = os.path.join(self.checkpoint_dir, "best.pth")
            shutil.copy(latest_path, best_path)

    def load_config(
        self, checkpoint: Mapping | bytes | str | os.PathLike, overwrite: bool = False, *args, **kwargs
    ) -> None:
        r"""
        Load config from checkpoint.

        Args:
            checkpoint: Checkpoint (or its path) to load.
            overwrite: If `True`, overwrite the current config with the loaded config.
                Defaults to `False`.
            *args: Additional arguments to pass to `self.load`.
            **kwargs: Additional keyword arguments to pass to `self.load`.

        Raises:
            FileNotFoundError: If `checkpoint` does not exists.
        """

        if isinstance(checkpoint, (bytes, str, os.PathLike)):
            if not os.path.exists(checkpoint):
                raise FileNotFoundError(f"checkpoint is set to {checkpoint!r} but does not exist")
            config = self.load(checkpoint, *args, **kwargs)
        elif isinstance(checkpoint, Mapping):
            config = checkpoint
        else:
            raise ValueError(f"checkpoint is set to {checkpoint!r} but is not a valid checkpoint")

        config = config.get("runner", config)
        self.config.merge(config, overwrite=overwrite)
        self.step_begin = config["steps"] + 1
        self.epoch_begin = config["epochs"] + 1

    def load_checkpoint(self, checkpoint: Mapping | bytes | str | os.PathLike, *args, **kwargs) -> None:
        """
        Restore model, optimizer, and scheduler states from a saved checkpoint.

        This method loads the complete training state from a checkpoint, including:
        - Model weights and parameters
        - Optimizer state (learning rates, momentum buffers, etc.)
        - Learning rate scheduler state

        The method supports loading from either:
        - A file path (str, bytes, PathLike)
        - An already loaded checkpoint dictionary (Mapping)

        Args:
            checkpoint: The checkpoint to load from, which can be:
                - A path to the checkpoint file
                - A pre-loaded checkpoint dictionary
            *args: Additional arguments passed to the underlying `self.load` method.
            **kwargs: Additional keyword arguments passed to the underlying `self.load` method.
                      Common options include `map_location` for controlling device mapping.

        Raises:
            ValueError: If `self.model` is not defined (must initialize model before loading).
            ValueError: If `checkpoint` is not a recognized checkpoint format.
            FileNotFoundError: If the checkpoint file specified does not exist.

        Note:
            - This method attempts to unwrap the model before loading (useful for distributed training wrappers)
            - Unlike `load_config`, this method loads both model parameters and optimizer/scheduler states
            - Missing optimizer or scheduler states in the checkpoint will produce warnings but not errors
            - The checkpoint is stored in `self.config.checkpoint` for reference

        Example:
            ```python
            # Load from a file
            runner.load_checkpoint('checkpoints/best.pth')

            # Load with device mapping
            runner.load_checkpoint('checkpoints/latest.pth', map_location='cuda:0')
            ```

        See Also:
            [`from_checkpoint`][danling.BaseRunner.from_checkpoint]: Build a new runner instance from checkpoint.
            [`load_pretrained`][danling.BaseRunner.load_pretrained]: Load only model parameters from a checkpoint.
            [`save_checkpoint`][danling.BaseRunner.save_checkpoint]: Save current state to a checkpoint.
        """

        if self.model is None:
            raise ValueError("model is not defined")
        if isinstance(checkpoint, (bytes, str, os.PathLike)):
            if not os.path.exists(checkpoint):
                raise FileNotFoundError(f"checkpoint is set to {checkpoint!r} but does not exist")
            ckpt = self.load(checkpoint, *args, **kwargs)
        elif isinstance(checkpoint, Mapping):
            ckpt = checkpoint
        else:
            raise ValueError(f"checkpoint is set to {checkpoint!r} but is not a valid checkpoint")

        state_dict = ckpt
        while "model" in state_dict or "module" in state_dict:
            state_dict = state_dict.get("model", state_dict)
            state_dict = state_dict.get("module", state_dict)
        self.unwrap(self.model).load_state_dict(state_dict)
        if self.optimizer is not None:
            if "optimizer" in ckpt:
                self.optimizer.load_state_dict(ckpt["optimizer"])
            else:
                warn("optimizer is not in checkpoint", category=RuntimeWarning, stacklevel=2)
        if self.scheduler is not None:
            if "scheduler" in ckpt:
                self.scheduler.load_state_dict(ckpt["scheduler"])
            else:
                warn("scheduler is not in checkpoint", category=RuntimeWarning, stacklevel=2)
        self.config.checkpoint = checkpoint

    @classmethod
    def from_checkpoint(cls, checkpoint: Mapping | bytes | str | os.PathLike, *args, **kwargs) -> BaseRunner:
        r"""
        Build BaseRunner from checkpoint.

        Args:
            checkpoint: Checkpoint (or its path) to load.
            *args: Additional arguments to pass to `cls.load`.
            **kwargs: Additional keyword arguments to pass to `cls.load`.

        Returns:
            (BaseRunner):
        """

        if isinstance(checkpoint, (bytes, str, os.PathLike)):
            ckpt = cls.load(checkpoint, *args, **kwargs)
        elif isinstance(checkpoint, Mapping):
            ckpt = checkpoint
        else:
            raise ValueError(f"checkpoint is set to {checkpoint} but is not a valid checkpoint")
        runner = cls(ckpt["runner"])
        runner.load_checkpoint(ckpt, override_config=False)
        return runner

    def load_pretrained(self, checkpoint: Mapping | bytes | str | os.PathLike, *args, **kwargs) -> None:
        """
        Load model from pretrained checkpoint.

        This method only loads the model weights.

        Args:
            checkpoint: Pretrained checkpoint (or its path) to load.
            *args: Additional arguments to pass to `self.load`.
            **kwargs: Additional keyword arguments to pass to `self.load`.

        Raises:
            ValueError: If `model` is not defined.
            ValueError: If `checkpoint` is not a valid checkpoint.
            FileNotFoundError: If `checkpoint` does not exists.

        See Also:
            [`load_checkpoint`][danling.BaseRunner.load_checkpoint]: Load model, optimizer, and scheduler from
                checkpoint.
        """

        if self.model is None:
            raise ValueError("model is not defined")
        if isinstance(checkpoint, (bytes, str, os.PathLike)):
            if not os.path.exists(checkpoint):
                raise FileNotFoundError(f"pretrained is set to {checkpoint!r} but does not exist")
            ckpt = self.load(checkpoint, *args, **kwargs)
        elif isinstance(checkpoint, Mapping):
            ckpt = checkpoint
        else:
            raise ValueError(f"pretrained is set to {checkpoint!r} but is not a valid checkpoint")

        state_dict = ckpt
        while "model" in state_dict or "module" in state_dict:
            state_dict = state_dict.get("model", state_dict)
            state_dict = state_dict.get("module", state_dict)
        self.unwrap(self.model).load_state_dict(state_dict)

    def get_step_result(self) -> NestedDict:
        result = self.meters.value()
        if self.metrics is not None:
            return self._merge_result(result, self.metrics.value())
        return result

    def get_epoch_result(self) -> NestedDict:
        result = self.meters.average()
        if self.metrics is not None:
            return self._merge_result(result, self.metrics.average())
        return result

    def _merge_result(self, meter_result, metric_result) -> NestedDict:
        for key, value in metric_result.items():
            if isinstance(value, (Mapping)) and len(value) == 1:
                value = next(iter(value.values()))
            metric_result[key] = value
        meter_result.update(metric_result)
        return meter_result

    def append_result(self, result: NestedDict, index: int | None = None) -> None:
        r"""
        Append result to `self.results`.

        Warnings:
            `self.results` is heavily relied upon for computing metrics.

            Failed to use this method may lead to unexpected behavior.
        """

        if index is None:
            index = self.epochs
            global __APPEND_RESULT_COUNTER__  # pylint: disable=global-statement
            __APPEND_RESULT_COUNTER__ += 1
            if index == 0 and __APPEND_RESULT_COUNTER__ > 1:
                warn(
                    """
                    Automatically set index to `self.epochs`.
                    Please ensure `self.epochs` updates before calling `append_result`
                    """,
                    category=RuntimeWarning,
                    stacklevel=2,
                )
        if index in self.results:
            self.results[index].merge(result)
        else:
            self.results[index] = result

    def print_result(self) -> None:
        r"""
        Print latest and best result.
        """

        print(f"latest result: {self.latest_result}")
        print(f"best result: {self.best_result}")

    def step_log(self, split: str, iteration: int, length: int | None = None):
        if length is None:
            length = len(self.dataloaders[split]) - 1
        result = self.get_step_result()
        print(self.format_step_result(result, split, iteration, length))
        if self.mode == "train":
            self.write_result(result, split)
        return result

    def format_step_result(
        self, result: NestedDict, split: str, steps: int, length: int, format_spec: str = ".4f"
    ) -> str:
        repr_str = ""
        if split is not None:
            if self.mode == "train":
                repr_str = f"training on {split} "
            elif self.mode == "evaluate":
                repr_str = f"evaluating on {split} "
            elif self.mode == "infer":
                repr_str = f"inferring on {split} "
            else:
                repr_str = f"running in {self.mode} mode on {split} "
        repr_str += f"[{steps}/{length}]\t"
        return repr_str + self.format_result(result, format_spec=format_spec)

    def format_epoch_result(
        self, result: NestedDict, epochs: int | None = None, epoch_end: int | None = None, format_spec: str = ".4f"
    ) -> str:
        epochs = epochs or self.epochs
        epoch_end = epoch_end or self.epoch_end
        repr_str = f"epoch [{epochs}/{epoch_end - 1}]" if epochs is not None and epoch_end else ""
        return repr_str + self.format_result(result, format_spec=format_spec)

    def format_result(self, result: Mapping, format_spec: str = ".4f") -> str:
        return format_result(result, format_spec=format_spec)

    def write_result(self, result: NestedDict, split: str, steps: int | None = None):
        if steps is None:
            steps = self.steps
        for name, score in result.all_items():
            name = name.replace(".", "/")
            if name == "loss" and isinstance(score, AverageMeter):
                score = score.avg
            if isinstance(score, Sequence):
                for i, s in enumerate(score):
                    self.write_score(f"{name}/{i}", s, split, steps)
            elif isinstance(score, Mapping):
                for k, s in score.items():
                    self.write_score(f"{name}/{k}", s, split, steps)
            else:
                self.write_score(name, score, split, steps)

    def write_score(self, name: str, score: float, split: str, steps: int):
        if self.writer:
            self.writer.add_scalar(f"{split}/{name}", score, steps)

    @catch
    @on_main_process
    def save_result(self) -> None:
        r"""
        Save result to `self.dir`.

        This method will save latest and best result to
        `self.dir/latest.json` and `self.dir/best.json` respectively.
        """

        results_path = os.path.join(self.dir, "results.json")
        self.save(
            {
                "name": self.name,
                "id": self.id,
                "timestamp": self.timestamp,
                "results": self.results,
            },
            results_path,
            indent=4,
        )
        ret = {"name": self.name, "id": self.id, "timestamp": self.timestamp}
        result = self.latest_result
        if isinstance(result, FlatDict):
            result = result.dict()
        # This is slower but ensure id is the first key
        if result is not None:
            ret.update(result)
        latest_path = os.path.join(self.dir, "latest.json")
        self.save(ret, latest_path, indent=4)
        if self.is_best:
            best_path = os.path.join(self.dir, "best.json")
            shutil.copy(latest_path, best_path)

    def unwrap(self, model: Any) -> Any:
        return model

    @cached_property
    def name(self) -> str:
        if "name" in self.config:
            return self.config["name"]
        return f"{self.config.experiment_name}-{self.config.run_name}"

    @cached_property
    def id(self) -> str:
        return f"{self.config.experiment_id:.8}{self.config.run_id:.8}"

    @cached_property
    def uuid(self) -> UUID:
        r"""
        UUID of the config.
        """

        return uuid5(self.run_uuid, self.id)

    @property
    def mode(self) -> RunnerMode:
        return self._mode

    @mode.setter
    def mode(self, mode: str | RunnerMode) -> None:
        if isinstance(mode, str):
            mode = RunnerMode(mode)
        self._mode = mode

    @property
    def config(self) -> Config:
        return self._config

    @property
    def batch_size(self) -> int:
        r"""
        Batch size.

        Note:
            If `self.dataloaders` is specified:
                - If `self.split` is specified, `batch_size` is the batch size of `self.split`.
                - If `self.train_splits` is specified, `batch_size` is the batch size of the first split.
                - If `self.evaluate_splits` is specified, `batch_size` is the batch size of the first split.
                - Otherwise, `batch_size` is the batch size of the first dataloader.
            Otherwise, `batch_size` is the batch size specified in `self.config.dataloader.batch_size`.
        """

        batch_size = None
        if self.dataloaders:
            if self.split:
                batch_size = self.dataloaders[self.split].batch_size
                if batch_size:
                    return batch_size
            train_splits = self.train_splits or []
            evaluate_splits = self.evaluate_splits or []
            splits = train_splits + evaluate_splits
            for split in splits:
                if split in self.dataloaders:
                    batch_size = self.dataloaders[split].batch_size
                    if batch_size:
                        return batch_size
            for dataloader in self.dataloaders.values():
                batch_size = dataloader.batch_size
                if batch_size:
                    return batch_size
        if not batch_size:
            batch_size = self.config.get("dataloader.batch_size")
        if batch_size:
            return batch_size
        raise AttributeError("batch_size could not be inferred and is not in config")

    @property
    def batch_size_equivalent(self) -> int:
        r"""
        The effective total batch size across all processes and gradient accumulation steps.

        This property calculates the total number of examples processed in a single
        optimization step by accounting for:
        - The base batch size per device
        - The number of processes (world_size for distributed training)
        - The number of gradient accumulation steps

        This is particularly useful for:
        - Learning rate scaling using the linear scaling rule
        - Calculating the total steps in an epoch
        - Estimating memory requirements across all devices

        Returns:
            (int): Effective batch size calculated as `batch_size * world_size * accum_steps`
        """

        return self.batch_size * self.world_size * self.accum_steps

    @property
    def epochs(self) -> int:
        r"""
        Current epoch.
        """
        return self.config.epochs

    @epochs.setter
    def epochs(self, epochs: int) -> None:
        self.config.epochs = epochs

    @property
    def epoch_begin(self) -> int:
        return self.config.epoch_begin

    @epoch_begin.setter
    def epoch_begin(self, epoch_begin: int) -> None:
        self.config.epoch_begin = epoch_begin

    @property
    def epoch_end(self) -> int | None:
        return self.config.epoch_end

    @epoch_end.setter
    def epoch_end(self, epoch_end: int) -> None:
        self.config.epoch_end = epoch_end

    @property
    def total_epochs(self) -> int:
        r"""
        Number of training epochs.
        """
        if self.epoch_end:
            return self.epoch_end - self.epoch_begin
        raise ValueError("epoch_end is not specified")

    @property
    def steps(self) -> int:
        r"""
        Current step.
        """
        return self.config.steps

    @steps.setter
    def steps(self, steps: int) -> None:
        self.config.steps = steps

    @property
    def step_begin(self) -> int:
        return self.config.step_begin

    @step_begin.setter
    def step_begin(self, step_begin: int) -> None:
        self.config.step_begin = step_begin

    @property
    def step_end(self) -> int | None:
        return self.config.step_end

    @step_end.setter
    def step_end(self, step_end: int) -> None:
        self.config.step_end = step_end

    @property
    def total_steps(self) -> int:
        r"""
        Number of training steps.
        """
        if self.step_end:
            return self.step_end - self.step_begin
        train_splits = self.train_splits
        if self.dataloaders and train_splits:
            return self.total_epochs * sum(len(self.dataloaders[split]) for split in train_splits)
        if self.datasets and train_splits:
            datasets_length = sum(len(self.datasets[split]) for split in self.train_splits)
            return self.total_epochs * ceil(datasets_length / self.batch_size / self.world_size)
        raise ValueError("total_steps could not be inferred and is not in config")

    @property
    def accum_steps(self) -> int:
        r"""
        Number of steps to accumulate gradients.
        """

        return self.config.get("accum_steps", 1)

    @property
    def progress(self) -> float:
        r"""
        Training Progress.

        Raises:
            RuntimeError: If no terminal is defined.
        """

        return self.steps / self.total_steps

    @property
    def train_splits(self) -> list[str]:
        if self._train_splits is not None:
            return self._train_splits
        if "train_splits" in self.config:
            self._train_splits = self.config["train_splits"]
            return self._train_splits
        if self.datasets:
            train_splits = [k for k, d in self.datasets.items() if getattr(d, "train", False)]
            if train_splits:
                self._train_splits = train_splits
                return train_splits
            dataset_splits = {k: getattr(d, "split", None) for k, d in self.datasets.items()}
            train_splits = [k for k, s in dataset_splits.items() if s == "train"]
            if train_splits:
                self._train_splits = train_splits
                return train_splits
        raise ValueError("train_splits could not be inferred and is not in config")

    @train_splits.setter
    def train_splits(self, train_splits: list[str]) -> None:
        self._train_splits = train_splits

    @property
    def evaluate_splits(self) -> list[str]:
        if self._evaluate_splits is not None:
            return self._evaluate_splits
        if "evaluate_splits" in self.config:
            self._evaluate_splits = self.config["evaluate_splits"]
            return self._evaluate_splits
        if self.datasets:
            evaluate_splits = [k for k in self.datasets if k not in self.train_splits]
            self._evaluate_splits = evaluate_splits
            return evaluate_splits
        raise ValueError("evaluate_splits could not be inferred and is not in config")

    @evaluate_splits.setter
    def evaluate_splits(self, evaluate_splits: list[str]) -> None:
        self._evaluate_splits = evaluate_splits

    @property
    def world_size(self) -> int:
        r"""
        Number of processes.
        """

        return 1

    @property
    def rank(self) -> int:
        r"""
        Process index of all processes.
        """

        return 0

    @property
    def local_rank(self) -> int:
        r"""
        Process index of local processes.
        """

        return 0

    @property
    def distributed(self) -> bool:
        r"""
        If runner is running in distributed mode.
        """

        return self.world_size > 1

    @property
    def is_main_process(self) -> bool:
        r"""
        If current process is the main process of all processes.
        """

        return self.rank == 0

    @property
    def is_local_main_process(self) -> bool:
        r"""
        If current process is the main process of local processes.
        """

        return self.local_rank == 0

    @property
    def best_fn(self) -> Callable:
        r"""
        Function to determine the best score from a list of scores.

        By default, the `best_fn` returns `min` if `self.config.score_name` is `loss`,
        otherwise, returns `max`.

        Subclass can override this method to accommodate needs, such as `min`.
        """

        return max if self.config.score_name != "loss" else min

    @property
    def best_index(self) -> int:
        r"""
        Find the best index from all scores.
        """

        if not self.scores:
            return 0
        return self.best_fn(reversed(self.scores), key=self.scores.get)

    @property
    def latest_result(self) -> NestedDict | None:
        r"""
        Latest result.
        """

        if not self.results:
            return None
        latest_index = next(reversed(self.results if PY38_PLUS else list(self.results)))  # type: ignore
        ret = self.results[latest_index].clone()
        ret["index"] = latest_index
        return ret

    @property
    def best_result(self) -> NestedDict | None:
        r"""
        Best result.
        """

        if not self.results:
            return None
        best_index = self.best_index
        ret = self.results[best_index].clone()
        ret["index"] = best_index
        return ret

    @property
    def scores(self) -> FlatDict | None:
        r"""
        All scores.

        Scores are extracted from results by `score_split` and `runner.config.score_name`,
        following `[r[score_split][self.config.score_name] for r in self.results]`.

        Scores are considered as the index of the performance of the model.
        It is useful to determine the best model and the best hyper-parameters.

        `score_split` is defined in `self.config.score_split`.
        If it is not set, `DanLing` will use `val` or `validate` if they appear in the `latest_result`.
        If `DanLing` still could not find, it will fall back to the second key in the `latest_result`
        if it contains more that one element, or the first key.

        Note that certain keys are ignored when falling back, they are defined in {defaults.IGNORED_NAMES_IN_METRICS}.
        """

        if not self.results:
            return None
        subsets = [i for i in self.latest_result.keys() if i not in defaults.IGNORED_NAMES_IN_METRICS]  # type: ignore
        score_split = self.config.get("score_split")
        if score_split is None and "val" in subsets:
            score_split = "val"
        if score_split is None and "validate" in subsets:
            score_split = "validate"
        if score_split is None:
            score_split = subsets[1] if len(subsets) > 1 else subsets[0]
        return FlatDict({k: v[score_split][self.config.score_name] for k, v in self.results.items()})

    @property
    def latest_score(self) -> float | None:
        r"""
        Latest score.
        """

        if not self.results:
            return None
        if not PY38_PLUS:
            return next(reversed(list(self.scores.values())))  # type: ignore
        return next(reversed(self.scores.values()))  # type: ignore

    @property
    def best_score(self) -> float | None:
        r"""
        Best score.
        """

        if not self.results:
            return None
        return self.scores[self.best_index]  # type: ignore

    @property
    def is_best(self) -> bool:
        r"""
        Determines whether the latest model checkpoint is the best performing one.

        This property compares the `latest_score` with the `best_score` to determine
        if the most recent model evaluation represents the best performance achieved
        so far. The comparison uses a small epsilon (1e-7) to handle floating point
        comparison issues.

        The property is used by:
        - `save_checkpoint` method to determine whether to save a "best.pth" checkpoint
        - Result logging and reporting mechanisms
        - Early stopping implementations

        Returns:
            (bool): True if the latest evaluation results represent the best performance
                 or if no results exist yet, False otherwise.

        Note:
            - Returns True if no results exist (first evaluation)
            - Uses numerical comparison with tolerance to avoid floating point issues
            - The definition of "best" depends on the metric (maximization for accuracy,
              minimization for loss, etc.) and is handled by the `best_fn` property

        See Also:
            [`best_score`][danling.BaseRunner.best_score]: The best score achieved so far.
            [`latest_score`][danling.BaseRunner.latest_score]: The most recent score.
            [`best_fn`][danling.BaseRunner.best_fn]: Function defining "best" (min/max).
        """

        if not self.results:
            return True
        try:
            return abs(self.latest_score - self.best_score) < 1e-7  # type: ignore
        except TypeError:
            return True

    @cached_ensure_dir
    def dir(self) -> str:
        r"""
        Directory of the run.
        """

        if "dir" in self.config:
            return self.config.dir
        return os.path.join(self.project_root, f"{self.name}-{self.id}", self.timestamp)

    @cached_ensure_parent_dir
    def log_file(self) -> str:
        r"""
        Path of log file.
        """

        if "log_file" in self.config:
            return self.config.log_file
        return os.path.join(self.dir, "run.log")

    @cached_ensure_dir
    def checkpoint_dir(self) -> str:
        r"""
        Directory of checkpoints.
        """

        if "checkpoint_dir" in self.config:
            return self.config.checkpoint_dir
        return os.path.join(self.dir, self.config.checkpoint_dir_name)

    # def __getattribute__(self, name) -> Any:
    #     if name in ("__class__", "__dict__"):
    #         return super().__getattribute__(name)
    #     if name in self.__dict__:
    #         return self.__dict__[name]
    #     if name in dir(self):
    #         return super().__getattribute__(name)
    #     if "config" in self and name in self.config:
    #         return self.config[name]
    #     return super().__getattribute__(name)

    def __getattr__(self, name) -> Any:
        if self.inited:
            if name in self.config:
                return self.config[name]
            if name in dir(self.config):
                return getattr(self.config, name)
        return super().__getattribute__(name)

    def __setattr__(self, name, value) -> None:
        if name in self.__dict__:
            if isinstance(self.__dict__[name], Variable):
                self.__dict__[name].set(value)
            else:
                self.__dict__[name] = value
            return
        if name in dir(self):
            if isinstance(super().__getattribute__(name), Variable):
                super().__getattribute__(name).set(value)
            else:
                object.__setattr__(self, name, value)
            return
        if self.inited:
            if name in self.config:
                if isinstance(self.config[name], Variable):
                    self.config[name].set(value)
                else:
                    self.config[name] = value
                return
            if name in dir(self.config):
                setattr(self.config, name, value)
                return
        object.__setattr__(self, name, value)

    def __contains__(self, name) -> bool:
        return name in dir(self) or ("config" in self.__dict__ and name in dir(self.config))

    def __repr__(self):
        lines = []
        for key, value in self.__dict__.items():
            if key.startswith("_") or key.endswith("_") or key == "inited":
                continue
            value_str = repr(value)
            value_str = self._add_indent(value_str)
            lines.append("(" + key + "): " + value_str)

        main_str = self.__class__.__name__ + "("
        if lines:
            main_str += "\n  " + "\n  ".join(lines) + "\n"

        main_str += ")"
        return main_str

    def _add_indent(self, text):
        lines = text.split("\n")
        # don't do anything for single-line stuff
        if len(lines) == 1:
            return text
        first = lines.pop(0)
        # add 2 spaces to each line but the first
        lines = [(2 * " ") + line for line in lines]
        lines = "\n".join(lines)
        lines = first + "\n" + lines
        return lines

batch_size property

Python
batch_size: int

Batch size.

Note

If self.dataloaders is specified: - If self.split is specified, batch_size is the batch size of self.split. - If self.train_splits is specified, batch_size is the batch size of the first split. - If self.evaluate_splits is specified, batch_size is the batch size of the first split. - Otherwise, batch_size is the batch size of the first dataloader. Otherwise, batch_size is the batch size specified in self.config.dataloader.batch_size.

batch_size_equivalent property

Python
batch_size_equivalent: int

The effective total batch size across all processes and gradient accumulation steps.

This property calculates the total number of examples processed in a single optimization step by accounting for: - The base batch size per device - The number of processes (world_size for distributed training) - The number of gradient accumulation steps

This is particularly useful for: - Learning rate scaling using the linear scaling rule - Calculating the total steps in an epoch - Estimating memory requirements across all devices

Returns:

Type Description
int

Effective batch size calculated as batch_size * world_size * accum_steps

epochs property writable

Python
epochs: int

Current epoch.

total_epochs property

Python
total_epochs: int

Number of training epochs.

steps property writable

Python
steps: int

Current step.

total_steps property

Python
total_steps: int

Number of training steps.

accum_steps property

Python
accum_steps: int

Number of steps to accumulate gradients.

progress property

Python
progress: float

Training Progress.

Raises:

Type Description
RuntimeError

If no terminal is defined.

world_size property

Python
world_size: int

Number of processes.

rank property

Python
rank: int

Process index of all processes.

local_rank property

Python
local_rank: int

Process index of local processes.

distributed property

Python
distributed: bool

If runner is running in distributed mode.

is_main_process property

Python
is_main_process: bool

If current process is the main process of all processes.

is_local_main_process property

Python
is_local_main_process: bool

If current process is the main process of local processes.

best_fn property

Python
best_fn: Callable

Function to determine the best score from a list of scores.

By default, the best_fn returns min if self.config.score_name is loss, otherwise, returns max.

Subclass can override this method to accommodate needs, such as min.

best_index property

Python
best_index: int

Find the best index from all scores.

latest_result property

Python
latest_result: NestedDict | None

Latest result.

best_result property

Python
best_result: NestedDict | None

Best result.

scores property

Python
scores: FlatDict | None

All scores.

Scores are extracted from results by score_split and runner.config.score_name, following [r[score_split][self.config.score_name] for r in self.results].

Scores are considered as the index of the performance of the model. It is useful to determine the best model and the best hyper-parameters.

score_split is defined in self.config.score_split. If it is not set, DanLing will use val or validate if they appear in the latest_result. If DanLing still could not find, it will fall back to the second key in the latest_result if it contains more that one element, or the first key.

Note that certain keys are ignored when falling back, they are defined in {defaults.IGNORED_NAMES_IN_METRICS}.

latest_score property

Python
latest_score: float | None

Latest score.

best_score property

Python
best_score: float | None

Best score.

is_best property

Python
is_best: bool

Determines whether the latest model checkpoint is the best performing one.

This property compares the latest_score with the best_score to determine if the most recent model evaluation represents the best performance achieved so far. The comparison uses a small epsilon (1e-7) to handle floating point comparison issues.

The property is used by: - save_checkpoint method to determine whether to save a “best.pth” checkpoint - Result logging and reporting mechanisms - Early stopping implementations

Returns:

Type Description
bool

True if the latest evaluation results represent the best performance or if no results exist yet, False otherwise.
Note
  • Returns True if no results exist (first evaluation)
  • Uses numerical comparison with tolerance to avoid floating point issues
  • The definition of “best” depends on the metric (maximization for accuracy, minimization for loss, etc.) and is handled by the best_fn property
See Also

best_score: The best score achieved so far. latest_score: The most recent score. best_fn: Function defining “best” (min/max).

init_distributed

Python
init_distributed() -> None

Initialise distributed running environment.

Source code in danling/runner/base_runner.py
Python
211
212
213
214
def init_distributed(self) -> None:
    r"""
    Initialise distributed running environment.
    """

init_logging

Python
init_logging() -> None

Set up logging.

Source code in danling/runner/base_runner.py
Python
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
@on_main_process
def init_logging(self) -> None:
    r"""
    Set up logging.
    """

    # Why is setting up proper logging so !@?#! ugly?
    logging.config.dictConfig(
        {
            "version": 1,
            "disable_existing_loggers": False,
            "formatters": {
                "standard": {"format": "%(asctime)s [%(levelname)s] %(name)s: %(message)s"},
            },
            "handlers": {
                "stdout": {
                    "level": "INFO",
                    "formatter": "standard",
                    "class": "logging.StreamHandler",
                    "stream": "ext://sys.stdout",
                },
                "logfile": {
                    "level": "DEBUG",
                    "formatter": "standard",
                    "class": "logging.FileHandler",
                    "filename": self.log_file,
                    "mode": "a",
                },
            },
            "loggers": {
                "": {
                    "handlers": ["stdout", "logfile"],
                    "level": "DEBUG",
                    "propagate": True,
                },
            },
        }
    )
    logging.captureWarnings(True)
    self.logger = logging.getLogger("runner")
    self.logger.flush = lambda: [h.flush() for h in self.logger.handlers]  # type: ignore

init_print

Python
init_print(process: int = 0) -> None

Set up print.

Only print on a specific process or when force = True.

Parameters:

Name Type Description Default
process
int

The process to print on.

 0
Note

If self.config.log = True, the default print function will be override by logging.info.

Source code in danling/runner/base_runner.py
Python
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
def init_print(self, process: int = 0) -> None:
    r"""
    Set up `print`.

    Only print on a specific `process` or when `force = True`.

    Args:
        process: The process to `print` on.

    Note:
        If `self.config.log = True`, the default `print` function will be override by `logging.info`.
    """

    logger = logging.getLogger("print")
    logger.flush = lambda: [h.flush for h in logger.handlers]  # type: ignore
    import builtins as __builtin__  # pylint: disable=C0415

    builtin_print = __builtin__.print

    @catch
    def print(*args, force=False, end="\n", file=None, flush=False, **kwargs):  # pylint: disable=redefined-builtin
        if self.rank == process or force:
            if self.config.log:
                if not args:
                    args = [""]
                logger.info(*args, **kwargs)
            else:
                builtin_print(*args, end=end, file=file, flush=flush, **kwargs)

    __builtin__.print = print

init_tensorboard

Python
init_tensorboard(*args, **kwargs) -> None

Set up Tensoraoard SummaryWriter.

Source code in danling/runner/base_runner.py
Python
289
290
291
292
293
294
@on_main_process
def init_tensorboard(self, *args, **kwargs) -> None:
    r"""
    Set up Tensoraoard SummaryWriter.
    """
    raise NotImplementedError

set_seed

Python
set_seed(seed: int = None, bias: int = None) -> int

Set random seeds for reproducibility across the application.

This method configures the random number generators for various libraries to ensure reproducible behavior in experiments. In the base implementation, it sets seeds for Python’s built-in random module and NumPy’s random generators if available.

Derived runners typically extend this method to set seeds for additional libraries (e.g., PyTorch, TensorFlow).

Parameters:

Name Type Description Default
seed
int

The base random seed to use. If None, uses the value from self.config.seed.

 None
bias
int

An offset to add to the seed for process-specific randomness. - If specified, adds this value to the seed - If None, uses self.rank (recommended for distributed training) - If False, disables bias (all processes use identical seed)

Using a per-process bias is important in distributed training to avoid all processes applying identical data augmentation.

 None

Returns:

Name Type Description
int int

The actual seed that was set (after applying bias)

Examples:

Set seed from config

runner.set_seed()

Set specific seed

runner.set_seed(42)

Set seed with explicit bias

runner.set_seed(42, bias=10)

Set same seed for all processes (not recommended for data augmentation)

runner.set_seed(42, bias=False)

Note
  • Setting seeds is essential for reproducible experiments
  • Different biases across processes ensure diverse data augmentation
  • Subclasses should call super().set_seed() if overriding
Source code in danling/runner/base_runner.py
Python
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
def set_seed(self, seed: int = None, bias: int = None) -> int:  # type: ignore[assignment]
    r"""
    Set random seeds for reproducibility across the application.

    This method configures the random number generators for various libraries
    to ensure reproducible behavior in experiments. In the base implementation,
    it sets seeds for Python's built-in `random` module and NumPy's random
    generators if available.

    Derived runners typically extend this method to set seeds for additional
    libraries (e.g., PyTorch, TensorFlow).

    Args:
        seed: The base random seed to use.
            If None, uses the value from `self.config.seed`.

        bias: An offset to add to the seed for process-specific randomness.
            - If specified, adds this value to the seed
            - If None, uses `self.rank` (recommended for distributed training)
            - If False, disables bias (all processes use identical seed)

            Using a per-process bias is important in distributed training to avoid
            all processes applying identical data augmentation.

    Returns:
        int: The actual seed that was set (after applying bias)

    Examples:
        # Set seed from config
        runner.set_seed()

        # Set specific seed
        runner.set_seed(42)

        # Set seed with explicit bias
        runner.set_seed(42, bias=10)

        # Set same seed for all processes (not recommended for data augmentation)
        runner.set_seed(42, bias=False)

    Note:
        - Setting seeds is essential for reproducible experiments
        - Different biases across processes ensure diverse data augmentation
        - Subclasses should call `super().set_seed()` if overriding
    """

    seed = seed or self.config.seed  # type: ignore[assignment]
    bias = bias or self.rank
    if bias:
        seed += bias
    if np_random is not None:
        np_random.seed(seed)
    random.seed(seed)
    return seed

set_deterministic

Python
set_deterministic() -> None

Set up deterministic.

Source code in danling/runner/base_runner.py
Python
351
352
353
354
355
356
def set_deterministic(self) -> None:
    r"""
    Set up deterministic.
    """

    raise NotImplementedError

scale_lr

Python
scale_lr(
    lr: float,
    lr_scale_factor: float | None = None,
    batch_size_base: int | None = None,
) -> float

Scale learning rate according to linear scaling rule.

Source code in danling/runner/base_runner.py
Python
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
def scale_lr(
    self,
    lr: float,
    lr_scale_factor: float | None = None,
    batch_size_base: int | None = None,
) -> float:
    r"""
    Scale learning rate according to [linear scaling rule](https://arxiv.org/abs/1706.02677).
    """

    if lr_scale_factor in self.config:
        lr_scale_factor = self.config.lr_scale_factor

    if lr_scale_factor is None:
        if batch_size_base is None:
            batch_size_base = getattr(self, "batch_size_base", None)
            if batch_size_base is None:
                raise ValueError("batch_size_base must be specified to auto scale lr")
        lr_scale_factor = self.batch_size_equivalent / batch_size_base
    elif batch_size_base is not None:
        warn(
            "batch_size_base will be ignored if lr_scale_factor is specified", category=RuntimeWarning, stacklevel=2
        )
    lr = lr * lr_scale_factor
    self.config.lr_scale_factor = lr_scale_factor
    return lr

advance

Python
advance(loss, *args, **kwargs) -> None

Backward loss and step optimizer & scheduler.

Parameters:

Name Type Description Default
loss

loss.

 required
Source code in danling/runner/base_runner.py
Python
385
386
387
388
389
390
391
392
393
def advance(self, loss, *args, **kwargs) -> None:
    r"""
    Backward loss and step optimizer & scheduler.

    Args:
        loss: loss.
    """

    raise NotImplementedError

state_dict

Python
state_dict(cls: Callable = dict) -> Mapping

Return dict of all attributes for checkpoint.

Source code in danling/runner/base_runner.py
Python
395
396
397
398
399
400
def state_dict(self, cls: Callable = dict) -> Mapping:
    r"""
    Return dict of all attributes for checkpoint.
    """

    return cls(self.config)

dict

Python
dict(cls: Callable = dict) -> Mapping

Convert config to Mapping.

Parameters:

Name Type Description Default
cls
Callable

Target `clc to convert to.

 dict
Source code in danling/runner/base_runner.py
Python
402
403
404
405
406
407
408
409
410
def dict(self, cls: Callable = dict) -> Mapping:
    r"""
    Convert config to Mapping.

    Args:
        cls: Target `clc to convert to.
    """

    return self.config.dict(cls)

save

Python
save(
    obj: Any,
    file: PathStr,
    main_process_only: bool = True,
    *args,
    **kwargs
) -> File

Save any file with supported extensions.

Runner.save internally calls dl.save, but with additional arguments to allow it save only on the main process. Moreover, any error raised by Runner.save will be caught and logged.

Source code in danling/runner/base_runner.py
Python
412
413
414
415
416
417
418
419
420
421
422
423
424
@catch
def save(self, obj: Any, file: PathStr, main_process_only: bool = True, *args, **kwargs) -> File:
    r"""
    Save any file with supported extensions.

    `Runner.save` internally calls `dl.save`,
    but with additional arguments to allow it save only on the main process.
    Moreover, any error raised by `Runner.save` will be caught and logged.
    """

    if main_process_only and self.is_main_process or not main_process_only:
        return save(obj, file, *args, **kwargs)
    return file

load staticmethod

Python
load(file: PathStr, *args, **kwargs) -> Any

Load any file with supported extensions.

Runner.load is identical to dl.load.

Source code in danling/runner/base_runner.py
Python
426
427
428
429
430
431
432
433
434
@staticmethod
def load(file: PathStr, *args, **kwargs) -> Any:
    r"""
    Load any file with supported extensions.

    `Runner.load` is identical to `dl.load`.
    """

    return load(file, *args, **kwargs)

json

Python
json(
    file: File,
    main_process_only: bool = True,
    *args,
    **kwargs
) -> None

Dump Runner config to json file.

Source code in danling/runner/base_runner.py
Python
436
437
438
439
440
441
442
443
@catch
def json(self, file: File, main_process_only: bool = True, *args, **kwargs) -> None:  # pylint: disable=R1710
    r"""
    Dump Runner config to json file.
    """

    if main_process_only and self.is_main_process or not main_process_only:
        return self.config.json(file, *args, **kwargs)

from_json classmethod

Python
from_json(file: File, *args, **kwargs) -> BaseRunner

Construct Runner from json file.

This function calls self.from_jsons() to construct object from json string. You may overwrite from_jsons in case something is not json serializable.

Source code in danling/runner/base_runner.py
Python
445
446
447
448
449
450
451
452
453
454
455
@classmethod
def from_json(cls, file: File, *args, **kwargs) -> BaseRunner:
    r"""
    Construct Runner from json file.

    This function calls `self.from_jsons()` to construct object from json string.
    You may overwrite `from_jsons` in case something is not json serializable.
    """

    with FlatDict.open(file) as fp:
        return cls.from_jsons(fp.read(), *args, **kwargs)

jsons

Python
jsons(*args, **kwargs) -> str

Dump Runner config to json string.

Source code in danling/runner/base_runner.py
Python
457
458
459
460
461
462
def jsons(self, *args, **kwargs) -> str:
    r"""
    Dump Runner config to json string.
    """

    return self.config.jsons(*args, **kwargs)

from_jsons classmethod

Python
from_jsons(string: str, *args, **kwargs) -> BaseRunner

Construct Runner from json string.

Source code in danling/runner/base_runner.py
Python
464
465
466
467
468
469
470
@classmethod
def from_jsons(cls, string: str, *args, **kwargs) -> BaseRunner:
    r"""
    Construct Runner from json string.
    """

    return cls(Config.from_jsons(string, *args, **kwargs))

yaml

Python
yaml(
    file: File,
    main_process_only: bool = True,
    *args,
    **kwargs
) -> None

Dump Runner config to yaml file.

Source code in danling/runner/base_runner.py
Python
472
473
474
475
476
477
478
479
@catch
def yaml(self, file: File, main_process_only: bool = True, *args, **kwargs) -> None:  # pylint: disable=R1710
    r"""
    Dump Runner config to yaml file.
    """

    if main_process_only and self.is_main_process or not main_process_only:
        return self.config.yaml(file, *args, **kwargs)

from_yaml classmethod

Python
from_yaml(file: File, *args, **kwargs) -> BaseRunner

Construct Runner from yaml file.

This function calls self.from_yamls() to construct object from yaml string. You may overwrite from_yamls in case something is not yaml serializable.

Source code in danling/runner/base_runner.py
Python
481
482
483
484
485
486
487
488
489
490
491
@classmethod
def from_yaml(cls, file: File, *args, **kwargs) -> BaseRunner:
    r"""
    Construct Runner from yaml file.

    This function calls `self.from_yamls()` to construct object from yaml string.
    You may overwrite `from_yamls` in case something is not yaml serializable.
    """

    with FlatDict.open(file) as fp:
        return cls.from_yamls(fp.read(), *args, **kwargs)

yamls

Python
yamls(*args, **kwargs) -> str

Dump Runner config to yaml string.

Source code in danling/runner/base_runner.py
Python
493
494
495
496
497
498
def yamls(self, *args, **kwargs) -> str:
    r"""
    Dump Runner config to yaml string.
    """

    return self.config.yamls(*args, **kwargs)

from_yamls classmethod

Python
from_yamls(string: str, *args, **kwargs) -> BaseRunner

Construct Runner from yaml string.

Source code in danling/runner/base_runner.py
Python
500
501
502
503
504
505
506
@classmethod
def from_yamls(cls, string: str, *args, **kwargs) -> BaseRunner:
    r"""
    Construct Runner from yaml string.
    """

    return cls(Config.from_yamls(string, *args, **kwargs))

check_dir

Python
check_dir(action: str = 'warn') -> bool

Check if self.dir is not empty.

Parameters:

Name Type Description Default
action
str

The action to perform if self.dir is not empty.

 'warn'
Source code in danling/runner/base_runner.py
Python
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
def check_dir(self, action: str = "warn") -> bool:
    r"""
    Check if `self.dir` is not empty.

    Args:
        action (str): The action to perform if `self.dir` is not empty.
        Can be one of ("warn", "raise", "ignore"), default is "warn".
    """

    if action and action not in ("warn", "raise", "ignore"):
        raise ValueError(f"action should be one of warn, raise or ignore, but got {action}")
    if os.listdir(self.dir):
        if action == "warn":
            warn(
                f"Directory `{self.dir}` is not empty",
                category=RuntimeWarning,
                stacklevel=2,
            )
        if action == "raise":
            raise RuntimeError(f"Directory `{self.dir}` is not empty")
        return False
    return True

save_checkpoint

Python
save_checkpoint(
    name: str = "latest",
    epochs: int | None = None,
    save_best: bool = True,
) -> None

Save the current model state and runner configuration to a checkpoint file.

This method saves the complete state of the runner, including model weights, optimizer state, scheduler state, and configuration to the checkpoint directory. The method handles various checkpoint naming strategies based on parameters.

Parameters:

Name Type Description Default
name
str

Base name of the checkpoint file (without extension). Defaults to "latest", which creates a checkpoint at {checkpoint_dir}/latest.pth.

 'latest'
epochs
int | None

Epoch number to record in the checkpoint. If None, uses self.epochs (current epoch counter).

 None
save_best
bool

Whether to also save a copy of the checkpoint as best.pth when the current model performance is the best so far (self.is_best is True). Defaults to True.

 True
Note
  • Checkpoints are saved only on the main process in distributed training.
  • The format is determined by the extension, defaulting to .pth.
  • The checkpoint includes the complete runner state from self.state_dict().
  • Periodic checkpoints are saved based on save_interval configuration: If self.config.save_interval > 0 and (epochs + 1) % save_interval == 0, the checkpoint will be copied to {checkpoint_dir}/epoch-{epochs}.pth.
See Also

load_checkpoint: Load a saved checkpoint. state_dict: Get the state dictionary to save.

Source code in danling/runner/base_runner.py
Python
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
@catch
@on_main_process
def save_checkpoint(self, name: str = "latest", epochs: int | None = None, save_best: bool = True) -> None:
    r"""
    Save the current model state and runner configuration to a checkpoint file.

    This method saves the complete state of the runner, including model weights,
    optimizer state, scheduler state, and configuration to the checkpoint directory.
    The method handles various checkpoint naming strategies based on parameters.

    Args:
        name: Base name of the checkpoint file (without extension).
            Defaults to `"latest"`, which creates a checkpoint at `{checkpoint_dir}/latest.pth`.
        epochs: Epoch number to record in the checkpoint.
            If None, uses `self.epochs` (current epoch counter).
        save_best: Whether to also save a copy of the checkpoint as `best.pth`
            when the current model performance is the best so far (`self.is_best` is True).
            Defaults to True.

    Note:
        - Checkpoints are saved only on the main process in distributed training.
        - The format is determined by the extension, defaulting to `.pth`.
        - The checkpoint includes the complete runner state from `self.state_dict()`.
        - Periodic checkpoints are saved based on `save_interval` configuration:
          If `self.config.save_interval > 0` and `(epochs + 1) % save_interval == 0`,
          the checkpoint will be copied to `{checkpoint_dir}/epoch-{epochs}.pth`.

    See Also:
        [`load_checkpoint`][danling.BaseRunner.load_checkpoint]: Load a saved checkpoint.
        [`state_dict`][danling.BaseRunner.state_dict]: Get the state dictionary to save.
    """

    epochs = epochs or self.epochs
    save_interval = self.config.get("save_interval", -1)
    latest_path = os.path.join(self.checkpoint_dir, f"{name}.pth")
    self.save(self.state_dict(), latest_path)
    if save_interval > 0 and (epochs + 1) % save_interval == 0:
        save_path = os.path.join(self.checkpoint_dir, f"epoch-{epochs}.pth")
        shutil.copy(latest_path, save_path)
    if save_best and self.is_best:
        best_path = os.path.join(self.checkpoint_dir, "best.pth")
        shutil.copy(latest_path, best_path)

load_config

Python
load_config(
    checkpoint: Mapping | bytes | str | PathLike,
    overwrite: bool = False,
    *args,
    **kwargs
) -> None

Load config from checkpoint.

Parameters:

Name Type Description Default
checkpoint
Mapping | bytes | str | PathLike

Checkpoint (or its path) to load.

 required
overwrite
bool

If True, overwrite the current config with the loaded config. Defaults to False.

 False
*args

Additional arguments to pass to self.load.

 ()
**kwargs

Additional keyword arguments to pass to self.load.

 {}

Raises:

Type Description
FileNotFoundError

If checkpoint does not exists.
Source code in danling/runner/base_runner.py
Python
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
def load_config(
    self, checkpoint: Mapping | bytes | str | os.PathLike, overwrite: bool = False, *args, **kwargs
) -> None:
    r"""
    Load config from checkpoint.

    Args:
        checkpoint: Checkpoint (or its path) to load.
        overwrite: If `True`, overwrite the current config with the loaded config.
            Defaults to `False`.
        *args: Additional arguments to pass to `self.load`.
        **kwargs: Additional keyword arguments to pass to `self.load`.

    Raises:
        FileNotFoundError: If `checkpoint` does not exists.
    """

    if isinstance(checkpoint, (bytes, str, os.PathLike)):
        if not os.path.exists(checkpoint):
            raise FileNotFoundError(f"checkpoint is set to {checkpoint!r} but does not exist")
        config = self.load(checkpoint, *args, **kwargs)
    elif isinstance(checkpoint, Mapping):
        config = checkpoint
    else:
        raise ValueError(f"checkpoint is set to {checkpoint!r} but is not a valid checkpoint")

    config = config.get("runner", config)
    self.config.merge(config, overwrite=overwrite)
    self.step_begin = config["steps"] + 1
    self.epoch_begin = config["epochs"] + 1

load_checkpoint

Python
load_checkpoint(
    checkpoint: Mapping | bytes | str | PathLike,
    *args,
    **kwargs
) -> None

Restore model, optimizer, and scheduler states from a saved checkpoint.

This method loads the complete training state from a checkpoint, including: - Model weights and parameters - Optimizer state (learning rates, momentum buffers, etc.) - Learning rate scheduler state

The method supports loading from either: - A file path (str, bytes, PathLike) - An already loaded checkpoint dictionary (Mapping)

Parameters:

Name Type Description Default
checkpoint
Mapping | bytes | str | PathLike

The checkpoint to load from, which can be: - A path to the checkpoint file - A pre-loaded checkpoint dictionary

 required
*args

Additional arguments passed to the underlying self.load method.

 ()
**kwargs

Additional keyword arguments passed to the underlying self.load method. Common options include map_location for controlling device mapping.

 {}

Raises:

Type Description
ValueError

If self.model is not defined (must initialize model before loading).
ValueError

If checkpoint is not a recognized checkpoint format.
FileNotFoundError

If the checkpoint file specified does not exist.
Note
  • This method attempts to unwrap the model before loading (useful for distributed training wrappers)
  • Unlike load_config, this method loads both model parameters and optimizer/scheduler states
  • Missing optimizer or scheduler states in the checkpoint will produce warnings but not errors
  • The checkpoint is stored in self.config.checkpoint for reference
Example
Python
1
2
3
4
5
# Load from a file
runner.load_checkpoint('checkpoints/best.pth')

# Load with device mapping
runner.load_checkpoint('checkpoints/latest.pth', map_location='cuda:0')
See Also

from_checkpoint: Build a new runner instance from checkpoint. load_pretrained: Load only model parameters from a checkpoint. save_checkpoint: Save current state to a checkpoint.

Source code in danling/runner/base_runner.py
Python
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
def load_checkpoint(self, checkpoint: Mapping | bytes | str | os.PathLike, *args, **kwargs) -> None:
    """
    Restore model, optimizer, and scheduler states from a saved checkpoint.

    This method loads the complete training state from a checkpoint, including:
    - Model weights and parameters
    - Optimizer state (learning rates, momentum buffers, etc.)
    - Learning rate scheduler state

    The method supports loading from either:
    - A file path (str, bytes, PathLike)
    - An already loaded checkpoint dictionary (Mapping)

    Args:
        checkpoint: The checkpoint to load from, which can be:
            - A path to the checkpoint file
            - A pre-loaded checkpoint dictionary
        *args: Additional arguments passed to the underlying `self.load` method.
        **kwargs: Additional keyword arguments passed to the underlying `self.load` method.
                  Common options include `map_location` for controlling device mapping.

    Raises:
        ValueError: If `self.model` is not defined (must initialize model before loading).
        ValueError: If `checkpoint` is not a recognized checkpoint format.
        FileNotFoundError: If the checkpoint file specified does not exist.

    Note:
        - This method attempts to unwrap the model before loading (useful for distributed training wrappers)
        - Unlike `load_config`, this method loads both model parameters and optimizer/scheduler states
        - Missing optimizer or scheduler states in the checkpoint will produce warnings but not errors
        - The checkpoint is stored in `self.config.checkpoint` for reference

    Example:
        ```python
        # Load from a file
        runner.load_checkpoint('checkpoints/best.pth')

        # Load with device mapping
        runner.load_checkpoint('checkpoints/latest.pth', map_location='cuda:0')
        ```

    See Also:
        [`from_checkpoint`][danling.BaseRunner.from_checkpoint]: Build a new runner instance from checkpoint.
        [`load_pretrained`][danling.BaseRunner.load_pretrained]: Load only model parameters from a checkpoint.
        [`save_checkpoint`][danling.BaseRunner.save_checkpoint]: Save current state to a checkpoint.
    """

    if self.model is None:
        raise ValueError("model is not defined")
    if isinstance(checkpoint, (bytes, str, os.PathLike)):
        if not os.path.exists(checkpoint):
            raise FileNotFoundError(f"checkpoint is set to {checkpoint!r} but does not exist")
        ckpt = self.load(checkpoint, *args, **kwargs)
    elif isinstance(checkpoint, Mapping):
        ckpt = checkpoint
    else:
        raise ValueError(f"checkpoint is set to {checkpoint!r} but is not a valid checkpoint")

    state_dict = ckpt
    while "model" in state_dict or "module" in state_dict:
        state_dict = state_dict.get("model", state_dict)
        state_dict = state_dict.get("module", state_dict)
    self.unwrap(self.model).load_state_dict(state_dict)
    if self.optimizer is not None:
        if "optimizer" in ckpt:
            self.optimizer.load_state_dict(ckpt["optimizer"])
        else:
            warn("optimizer is not in checkpoint", category=RuntimeWarning, stacklevel=2)
    if self.scheduler is not None:
        if "scheduler" in ckpt:
            self.scheduler.load_state_dict(ckpt["scheduler"])
        else:
            warn("scheduler is not in checkpoint", category=RuntimeWarning, stacklevel=2)
    self.config.checkpoint = checkpoint

from_checkpoint classmethod

Python
from_checkpoint(
    checkpoint: Mapping | bytes | str | PathLike,
    *args,
    **kwargs
) -> BaseRunner

Build BaseRunner from checkpoint.

Parameters:

Name Type Description Default
checkpoint
Mapping | bytes | str | PathLike

Checkpoint (or its path) to load.

 required
*args

Additional arguments to pass to cls.load.

 ()
**kwargs

Additional keyword arguments to pass to cls.load.

 {}

Returns:

Type Description
BaseRunner
Source code in danling/runner/base_runner.py
Python
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
@classmethod
def from_checkpoint(cls, checkpoint: Mapping | bytes | str | os.PathLike, *args, **kwargs) -> BaseRunner:
    r"""
    Build BaseRunner from checkpoint.

    Args:
        checkpoint: Checkpoint (or its path) to load.
        *args: Additional arguments to pass to `cls.load`.
        **kwargs: Additional keyword arguments to pass to `cls.load`.

    Returns:
        (BaseRunner):
    """

    if isinstance(checkpoint, (bytes, str, os.PathLike)):
        ckpt = cls.load(checkpoint, *args, **kwargs)
    elif isinstance(checkpoint, Mapping):
        ckpt = checkpoint
    else:
        raise ValueError(f"checkpoint is set to {checkpoint} but is not a valid checkpoint")
    runner = cls(ckpt["runner"])
    runner.load_checkpoint(ckpt, override_config=False)
    return runner

load_pretrained

Python
load_pretrained(
    checkpoint: Mapping | bytes | str | PathLike,
    *args,
    **kwargs
) -> None

Load model from pretrained checkpoint.

This method only loads the model weights.

Parameters:

Name Type Description Default
checkpoint
Mapping | bytes | str | PathLike

Pretrained checkpoint (or its path) to load.

 required
*args

Additional arguments to pass to self.load.

 ()
**kwargs

Additional keyword arguments to pass to self.load.

 {}

Raises:

Type Description
ValueError

If model is not defined.
ValueError

If checkpoint is not a valid checkpoint.
FileNotFoundError

If checkpoint does not exists.
See Also

load_checkpoint: Load model, optimizer, and scheduler from checkpoint.

Source code in danling/runner/base_runner.py
Python
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
def load_pretrained(self, checkpoint: Mapping | bytes | str | os.PathLike, *args, **kwargs) -> None:
    """
    Load model from pretrained checkpoint.

    This method only loads the model weights.

    Args:
        checkpoint: Pretrained checkpoint (or its path) to load.
        *args: Additional arguments to pass to `self.load`.
        **kwargs: Additional keyword arguments to pass to `self.load`.

    Raises:
        ValueError: If `model` is not defined.
        ValueError: If `checkpoint` is not a valid checkpoint.
        FileNotFoundError: If `checkpoint` does not exists.

    See Also:
        [`load_checkpoint`][danling.BaseRunner.load_checkpoint]: Load model, optimizer, and scheduler from
            checkpoint.
    """

    if self.model is None:
        raise ValueError("model is not defined")
    if isinstance(checkpoint, (bytes, str, os.PathLike)):
        if not os.path.exists(checkpoint):
            raise FileNotFoundError(f"pretrained is set to {checkpoint!r} but does not exist")
        ckpt = self.load(checkpoint, *args, **kwargs)
    elif isinstance(checkpoint, Mapping):
        ckpt = checkpoint
    else:
        raise ValueError(f"pretrained is set to {checkpoint!r} but is not a valid checkpoint")

    state_dict = ckpt
    while "model" in state_dict or "module" in state_dict:
        state_dict = state_dict.get("model", state_dict)
        state_dict = state_dict.get("module", state_dict)
    self.unwrap(self.model).load_state_dict(state_dict)

append_result

Python
append_result(
    result: NestedDict, index: int | None = None
) -> None

Append result to self.results.

Source code in danling/runner/base_runner.py
Python
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
def append_result(self, result: NestedDict, index: int | None = None) -> None:
    r"""
    Append result to `self.results`.

    Warnings:
        `self.results` is heavily relied upon for computing metrics.

        Failed to use this method may lead to unexpected behavior.
    """

    if index is None:
        index = self.epochs
        global __APPEND_RESULT_COUNTER__  # pylint: disable=global-statement
        __APPEND_RESULT_COUNTER__ += 1
        if index == 0 and __APPEND_RESULT_COUNTER__ > 1:
            warn(
                """
                Automatically set index to `self.epochs`.
                Please ensure `self.epochs` updates before calling `append_result`
                """,
                category=RuntimeWarning,
                stacklevel=2,
            )
    if index in self.results:
        self.results[index].merge(result)
    else:
        self.results[index] = result

print_result

Python
print_result() -> None

Print latest and best result.

Source code in danling/runner/base_runner.py
Python
790
791
792
793
794
795
796
def print_result(self) -> None:
    r"""
    Print latest and best result.
    """

    print(f"latest result: {self.latest_result}")
    print(f"best result: {self.best_result}")

save_result

Python
save_result() -> None

Save result to self.dir.

This method will save latest and best result to self.dir/latest.json and self.dir/best.json respectively.

Source code in danling/runner/base_runner.py
Python
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
@catch
@on_main_process
def save_result(self) -> None:
    r"""
    Save result to `self.dir`.

    This method will save latest and best result to
    `self.dir/latest.json` and `self.dir/best.json` respectively.
    """

    results_path = os.path.join(self.dir, "results.json")
    self.save(
        {
            "name": self.name,
            "id": self.id,
            "timestamp": self.timestamp,
            "results": self.results,
        },
        results_path,
        indent=4,
    )
    ret = {"name": self.name, "id": self.id, "timestamp": self.timestamp}
    result = self.latest_result
    if isinstance(result, FlatDict):
        result = result.dict()
    # This is slower but ensure id is the first key
    if result is not None:
        ret.update(result)
    latest_path = os.path.join(self.dir, "latest.json")
    self.save(ret, latest_path, indent=4)
    if self.is_best:
        best_path = os.path.join(self.dir, "best.json")
        shutil.copy(latest_path, best_path)

uuid

Python
uuid() -> UUID

UUID of the config.

Source code in danling/runner/base_runner.py
Python
901
902
903
904
905
906
907
@cached_property
def uuid(self) -> UUID:
    r"""
    UUID of the config.
    """

    return uuid5(self.run_uuid, self.id)

dir

Python
dir() -> str

Directory of the run.

Source code in danling/runner/base_runner.py
Python
1308
1309
1310
1311
1312
1313
1314
1315
1316
@cached_ensure_dir
def dir(self) -> str:
    r"""
    Directory of the run.
    """

    if "dir" in self.config:
        return self.config.dir
    return os.path.join(self.project_root, f"{self.name}-{self.id}", self.timestamp)

log_file

Python
log_file() -> str

Path of log file.

Source code in danling/runner/base_runner.py
Python
1318
1319
1320
1321
1322
1323
1324
1325
1326
@cached_ensure_parent_dir
def log_file(self) -> str:
    r"""
    Path of log file.
    """

    if "log_file" in self.config:
        return self.config.log_file
    return os.path.join(self.dir, "run.log")

checkpoint_dir

Python
checkpoint_dir() -> str

Directory of checkpoints.

Source code in danling/runner/base_runner.py
Python
1328
1329
1330
1331
1332
1333
1334
1335
1336
@cached_ensure_dir
def checkpoint_dir(self) -> str:
    r"""
    Directory of checkpoints.
    """

    if "checkpoint_dir" in self.config:
        return self.config.checkpoint_dir
    return os.path.join(self.dir, self.config.checkpoint_dir_name)

Config

Bases: Config

Config is a Config that contains all states of a Runner.

Config is designed to store all critical information of a Run so that you can resume a run from a state and corresponding weights or even restart a run from a state.

Config is also designed to be serialisable and hashable, so that you can save it to a file. Config is saved in checkpoint together with weights by default.

Since Config is a Config, you can access its attributes by state["key"] or state.key.

General:

Name Type Description
run_name str

Defaults to "DanLing".
run_id str

hex of self.run_uuid.
run_uuid (UUID, property)

uuid5(self.experiment_id, str(hash(self))).
experiment_name str

Defaults to "DanLing".
experiment_id str

git hash of the current HEAD. Defaults to "xxxxxxxxxxxxxxxx" if Runner not under a git repo or git/gitpython not installed.
experiment_uuid (UUID, property)

UUID of self.experiment_id. Defaults to UUID('78787878-7878-7878-7878-787878787878') if Runner not under a git repo or git/gitpython not installed.

Reproducibility:

Name Type Description
seed int

Defaults to randint(0, 2**32 - 1).
deterministic bool

Ensure deterministic operations. Defaults to False.

Progress:

Name Type Description
steps int

The number of steps calls.
epochs int

The number of complete passes over the datasets.
step_end int

End running step. Note that step_end not initialised since this variable may not apply to some Runners.
epoch_end int

End running epoch. Note that epoch_end not initialised since this variable may not apply to some Runners.

In general you should use either step_end or epoch_end to indicate the length of running.

IO:

Name Type Description
project_root str

The root directory for all experiments. Defaults to "experiments".

project_root is the root directory of all Experiments, and should be consistent across the Project.

dir is the directory of a certain Run.

There is no attributes/properties for Group and Experiment.

checkpoint_dir_name is relative to dir, and is passed to generate checkpoint_dir (checkpoint_dir = os.path.join(dir, checkpoint_dir_name)). In practice, checkpoint_dir_name is rarely called.

logging:

Name Type Description
log bool

Whether to log the outputs. Defaults to True.
tensorboard bool

Whether to use tensorboard. Defaults to False.
log_interval int

Interval of printing logs. Defaults to None, print logs every 1/10 of the longest split.
save_interval int

Interval of saving intermediate checkpoints. Defaults to None, never save checkpoints. If <= 0, save only the latest and the best checkpoints.
Note

Config is a Config, so you can access its attributes by state["name"] or state.name.

See Also

BaseRunner: The base runner class.

Source code in danling/runner/config.py
Python
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
class Config(chanfig.Config):  # pylint: disable=too-many-instance-attributes
    r"""
    `Config` is a [`Config`][chanfig.Config] that contains all states of a `Runner`.

    `Config` is designed to store all critical information of a Run so that you can resume a run
    from a state and corresponding weights or even restart a run from a state.

    `Config` is also designed to be serialisable and hashable, so that you can save it to a file.
    `Config` is saved in checkpoint together with weights by default.

    Since `Config` is a [`Config`][chanfig.Config], you can access its attributes by
    `state["key"]` or `state.key`.

    Attributes: General:
        run_name (str): Defaults to `"DanLing"`.
        run_id (str): hex of `self.run_uuid`.
        run_uuid (UUID, property): `uuid5(self.experiment_id, str(hash(self)))`.
        experiment_name (str): Defaults to `"DanLing"`.
        experiment_id (str): git hash of the current HEAD.
            Defaults to `"xxxxxxxxxxxxxxxx"` if Runner not under a git repo or git/gitpython not installed.
        experiment_uuid (UUID, property): UUID of `self.experiment_id`.
            Defaults to `UUID('78787878-7878-7878-7878-787878787878')`
            if Runner not under a git repo or git/gitpython not installed.

    Attributes: Reproducibility:
        seed (int): Defaults to `randint(0, 2**32 - 1)`.
        deterministic (bool): Ensure [deterministic](https://pytorch.org/docs/stable/notes/randomness.html) operations.
            Defaults to `False`.

    Attributes: Progress:
        steps (int): The number of `steps` calls.
        epochs (int): The number of complete passes over the datasets.
        step_end (int): End running step.
            Note that `step_end` not initialised since this variable may not apply to some Runners.
        epoch_end (int): End running epoch.
            Note that `epoch_end` not initialised since this variable may not apply to some Runners.

    In general you should use either `step_end` or `epoch_end` to indicate the length of running.

    Attributes: IO:
        project_root (str): The root directory for all experiments.
            Defaults to `"experiments"`.

    `project_root` is the root directory of all **Experiments**, and should be consistent across the **Project**.

    `dir` is the directory of a certain **Run**.

    There is no attributes/properties for **Group** and **Experiment**.

    `checkpoint_dir_name` is relative to `dir`, and is passed to generate `checkpoint_dir`
    (`checkpoint_dir = os.path.join(dir, checkpoint_dir_name)`).
    In practice, `checkpoint_dir_name` is rarely called.

    Attributes: logging:
        log (bool): Whether to log the outputs.
            Defaults to `True`.
        tensorboard (bool): Whether to use `tensorboard`.
            Defaults to `False`.
        log_interval (int): Interval of printing logs.
            Defaults to `None`, print logs every 1/10 of the longest split.
        save_interval (int): Interval of saving intermediate checkpoints.
            Defaults to `None`, never save checkpoints.
            If <= 0, save only the latest and the best checkpoints.

    Note:
        `Config` is a [`Config`][chanfig.Config], so you can access its attributes by `state["name"]` or `state.name`.

    See Also:
        [`BaseRunner`][danling.runner.BaseRunner]: The base runner class.
    """

    # DO NOT set default value in class, as they won't be stored in `__dict__`.

    run_name: str = defaults.RUN_NAME
    run_id: str
    experiment_name: str = defaults.EXPERIMENT_NAME
    experiment_id: str

    seed: Optional[int] = None
    deterministic: bool = False

    steps: int = 0
    epochs: int = 0
    step_begin: int = 0
    epoch_begin: int = 0
    step_end: Optional[int] = None
    epoch_end: Optional[int] = None

    score_split: Optional[str] = None
    score_name: str = "loss"

    project_root: str = "experiments"
    checkpoint_dir_name: str = "checkpoints"
    log: bool = True
    tensorboard: bool = False
    log_interval: Optional[int] = None
    save_interval: Optional[int] = None

    def __post_init__(self):
        if "experiment_id" not in self:
            self.experiment_id = get_git_hash() or defaults.EXPERIMENT_ID
        if "run_id" not in self:
            self.run_id = self.run_uuid.hex
        self.setattr("ignored_keys_in_hash", defaults.IGNORED_NAMES_IN_HASH)

    @property
    def experiment_uuid(self) -> UUID:
        r"""
        UUID of the experiment.
        """

        return UUID(bytes=bytes(self.experiment_id.ljust(16, "x")[:16], encoding="ascii"))

    @property
    def run_uuid(self) -> UUID:
        r"""
        UUID of the run.
        """

        ignored_keys_in_hash = self.getattr("ignored_keys_in_hash", defaults.IGNORED_NAMES_IN_HASH)
        state: chanfig.Config = chanfig.Config({k: v for k, v in self.dict().items() if k not in ignored_keys_in_hash})
        return uuid5(self.experiment_uuid, state.yamls())

    def __hash__(self) -> int:
        return int(self.run_uuid.hex, 16)

experiment_uuid property

Python
experiment_uuid: UUID

UUID of the experiment.

run_uuid property

Python
run_uuid: UUID

UUID of the run.

DeepSpeedRunner

Bases: TorchRunner

Set up everything for running a job with DeepSpeed.

DeepSpeed is a distributed training framework that provides a more efficient way to run large-scale models.

Configure DeepSpeedRunner is tough, but once you get the hang of it, it’s a powerful tool. Read more about DeepSpeed at DeepSpeed’s official website.

Source code in danling/runner/deepspeed_runner.py
Python
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
class DeepSpeedRunner(TorchRunner):
    r"""
    Set up everything for running a job with DeepSpeed.

    DeepSpeed is a distributed training framework that provides a more efficient way to run large-scale models.

    Configure DeepSpeedRunner is tough, but once you get the hang of it, it's a powerful tool.
    Read more about DeepSpeed at [DeepSpeed's official website](https://www.deepspeed.ai/).
    """

    def __init__(self, config: Config) -> None:
        ds.check()
        super().__init__(config)

    def init_distributed(self) -> None:
        r"""
        Set up distributed training.

        Initialise process group and set up DDP variables.
        """

        backend = self.config.get("backend", os.getenv("BACKEND"))
        init_method = self.config.get("init_method", os.getenv("INIT_METHOD"))
        world_size = int(self.config.get("world_size", os.getenv("WORLD_SIZE", "1")))
        rank = int(self.config.get("rank", os.getenv("RANK", "0")))
        if world_size > 1:
            if torch.cuda.is_available():
                torch.cuda.set_device(self.get_local_rank())
            deepspeed.init_distributed(dist_backend=backend, init_method=init_method, world_size=world_size, rank=rank)
            object_list = [self.id, self.timestamp]
            dist.broadcast_object_list(object_list)
            self.id, self.timestamp = object_list

    def __post_init__(self):
        super().__post_init__()
        self.config.deepspeed = self.get_deepspeed_config()
        self.model, self.optimizer, _, self.scheduler = deepspeed.initialize(
            model=self.model,
            optimizer=self.optimizer,
            lr_scheduler=self.scheduler,
            config=self.config.deepspeed,
        )

    def advance(self, loss) -> None:
        self.backward(loss)
        if self.config.get("max_grad_value") is not None:
            clip_grad_value_(self.model.parameters(), self.config["max_grad_value"])
        self.model.step()
        if self.ema is not None:
            self.ema.update()
        self.config.steps = self.model.global_steps

    def backward(self, loss: torch.Tensor) -> None:
        return self.model.backward(loss)

    def get_local_rank(self) -> int:
        local_rank = self.config.get("local_rank", os.getenv("LOCAL_RANK"))
        if local_rank is not None:
            return int(local_rank)
        rank = self.config.get("rank", os.getenv("RANK"))
        world_size = self.config.get("world_size", os.getenv("WORLD_SIZE"))
        if world_size is None or rank is None:
            raise ValueError("Please provide either `local_rank` or `world_size` and `rank`")
        return int(world_size) % int(rank)

    def unwrap(self, model: nn.Module) -> nn.Module:
        while isinstance(model, (deepspeed.DeepSpeedEngine, nn.parallel.DistributedDataParallel)):
            model = model.module
        return model

    @property
    def deepspeed(self) -> NestedDict | None:
        if isinstance(self.model, deepspeed.DeepSpeedEngine):
            return self.model.config
        return None

    @catch
    def save_checkpoint(self, name: str = "latest", epoch: int | None = None, save_best: bool = True) -> None:
        r"""
        Save checkpoint to `self.checkpoint_dir`.

        Args:
            name: Name of the checkpoint. Defaults to `"latest"`.
            epoch: Epoch to save. Defaults to `self.epochs`.
            save_best: If `True`, when `self.is_best` is `True`, the checkpoint will also be copied to
                `self.checkpoint_dir/best`.

        If `self.config.save_interval` is positive and `epochs + 1` is a multiple of `save_interval`,
        the checkpoint will also be copied to `self.checkpoint_dir/epoch-{epochs}`.
        """

        epoch = epoch or self.epochs
        save_interval = self.config.get("save_interval", -1)
        latest_path = os.path.join(self.checkpoint_dir, name)
        os.makedirs(latest_path, exist_ok=True)
        self.yaml(os.path.join(latest_path, "runner.yaml"))
        self.model.save_checkpoint(
            self.checkpoint_dir, tag=name, client_state={"runner": self.config.dict()}, save_latest=False
        )
        if save_interval > 0 and (epoch + 1) % save_interval == 0:
            save_path = os.path.join(self.checkpoint_dir, f"epoch-{epoch}")
            shutil.copytree(latest_path, save_path, dirs_exist_ok=True)
        if save_best and self.is_best:
            best_path = os.path.join(self.checkpoint_dir, "best")
            shutil.copytree(latest_path, best_path, dirs_exist_ok=True)

    def load_checkpoint(self, checkpoint: bytes | str | os.PathLike, *args, **kwargs) -> None:  # type: ignore[override]
        """
        Load model, optimizer, and scheduler from checkpoint.

        Args:
            checkpoint: Checkpoint (or its path) to load.
            *args: Additional arguments to pass to `self.load`.
            **kwargs: Additional keyword arguments to pass to `self.load`.

        Raises:
            ValueError: If `model` is not defined.
            ValueError: If `model` is not an instance of `deepspeed.DeepSpeedEngine`.

        See Also:
            [`from_checkpoint`][danling.BaseRunner.from_checkpoint]: Build runner from checkpoint.
            [`load_pretrained`][danling.BaseRunner.load_pretrained]: Load model parameters from pretrained checkpoint.
        """

        if self.model is None:
            raise ValueError("model is not defined")
        if not isinstance(self.model, deepspeed.DeepSpeedEngine):
            raise ValueError("model is not an instance of `deepspeed.DeepSpeedEngine`")

        self.model.load_checkpoint(checkpoint)
        self.config.checkpoint = checkpoint

    def load_pretrained(self, checkpoint: bytes | str | os.PathLike, *args, **kwargs) -> None:  # type: ignore[override]
        """
        Load model from pretrained checkpoint.

        This method only loads the model weights.

        Args:
            checkpoint: Pretrained checkpoint directory.
            *args: Additional arguments to pass to `self.load`.
            **kwargs: Additional keyword arguments to pass to `self.load`.

        Raises:
            ValueError: If `model` is not defined.

        See Also:
            [`load_checkpoint`][danling.BaseRunner.load_checkpoint]: Load model, optimizer, and scheduler from
                checkpoint.
        """

        if self.model is None:
            raise ValueError("model is not defined")

        self.model.load_checkpoint(checkpoint, load_module_only=True)
        self.config.pretrained = checkpoint

    def load_config(
        self, checkpoint: bytes | str | os.PathLike, overwrite: bool = False, *args, **kwargs  # type: ignore[override]
    ) -> None:
        r"""
        Load config from checkpoint.

        Args:
            checkpoint: Checkpoint (or its path) to load.
            overwrite: If `True`, overwrite the current config with the loaded config.
                Defaults to `False`.
            *args: Additional arguments to pass to `self.load`.
            **kwargs: Additional keyword arguments to pass to `self.load`.

        Raises:
            FileNotFoundError: If `checkpoint` does not exists.
        """

        if isinstance(checkpoint, bytes):
            checkpoint = checkpoint.decode()

        config = self.load(os.path.join(checkpoint, "runner.yaml"), *args, **kwargs)
        self.config.merge(config, overwrite=overwrite)
        self.step_begin = config["steps"] + 1
        self.epoch_begin = config["epochs"] + 1

init_distributed

Python
init_distributed() -> None

Set up distributed training.

Initialise process group and set up DDP variables.

Source code in danling/runner/deepspeed_runner.py
Python
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
def init_distributed(self) -> None:
    r"""
    Set up distributed training.

    Initialise process group and set up DDP variables.
    """

    backend = self.config.get("backend", os.getenv("BACKEND"))
    init_method = self.config.get("init_method", os.getenv("INIT_METHOD"))
    world_size = int(self.config.get("world_size", os.getenv("WORLD_SIZE", "1")))
    rank = int(self.config.get("rank", os.getenv("RANK", "0")))
    if world_size > 1:
        if torch.cuda.is_available():
            torch.cuda.set_device(self.get_local_rank())
        deepspeed.init_distributed(dist_backend=backend, init_method=init_method, world_size=world_size, rank=rank)
        object_list = [self.id, self.timestamp]
        dist.broadcast_object_list(object_list)
        self.id, self.timestamp = object_list

save_checkpoint

Python
save_checkpoint(
    name: str = "latest",
    epoch: int | None = None,
    save_best: bool = True,
) -> None

Save checkpoint to self.checkpoint_dir.

Parameters:

Name Type Description Default
name
str

Name of the checkpoint. Defaults to "latest".

 'latest'
epoch
int | None

Epoch to save. Defaults to self.epochs.

 None
save_best
bool

If True, when self.is_best is True, the checkpoint will also be copied to self.checkpoint_dir/best.

 True

If self.config.save_interval is positive and epochs + 1 is a multiple of save_interval, the checkpoint will also be copied to self.checkpoint_dir/epoch-{epochs}.

Source code in danling/runner/deepspeed_runner.py
Python
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
@catch
def save_checkpoint(self, name: str = "latest", epoch: int | None = None, save_best: bool = True) -> None:
    r"""
    Save checkpoint to `self.checkpoint_dir`.

    Args:
        name: Name of the checkpoint. Defaults to `"latest"`.
        epoch: Epoch to save. Defaults to `self.epochs`.
        save_best: If `True`, when `self.is_best` is `True`, the checkpoint will also be copied to
            `self.checkpoint_dir/best`.

    If `self.config.save_interval` is positive and `epochs + 1` is a multiple of `save_interval`,
    the checkpoint will also be copied to `self.checkpoint_dir/epoch-{epochs}`.
    """

    epoch = epoch or self.epochs
    save_interval = self.config.get("save_interval", -1)
    latest_path = os.path.join(self.checkpoint_dir, name)
    os.makedirs(latest_path, exist_ok=True)
    self.yaml(os.path.join(latest_path, "runner.yaml"))
    self.model.save_checkpoint(
        self.checkpoint_dir, tag=name, client_state={"runner": self.config.dict()}, save_latest=False
    )
    if save_interval > 0 and (epoch + 1) % save_interval == 0:
        save_path = os.path.join(self.checkpoint_dir, f"epoch-{epoch}")
        shutil.copytree(latest_path, save_path, dirs_exist_ok=True)
    if save_best and self.is_best:
        best_path = os.path.join(self.checkpoint_dir, "best")
        shutil.copytree(latest_path, best_path, dirs_exist_ok=True)

load_checkpoint

Python
load_checkpoint(
    checkpoint: bytes | str | PathLike, *args, **kwargs
) -> None

Load model, optimizer, and scheduler from checkpoint.

Parameters:

Name Type Description Default
checkpoint
bytes | str | PathLike

Checkpoint (or its path) to load.

 required
*args

Additional arguments to pass to self.load.

 ()
**kwargs

Additional keyword arguments to pass to self.load.

 {}

Raises:

Type Description
ValueError

If model is not defined.
ValueError

If model is not an instance of deepspeed.DeepSpeedEngine.
See Also

from_checkpoint: Build runner from checkpoint. load_pretrained: Load model parameters from pretrained checkpoint.

Source code in danling/runner/deepspeed_runner.py
Python
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
def load_checkpoint(self, checkpoint: bytes | str | os.PathLike, *args, **kwargs) -> None:  # type: ignore[override]
    """
    Load model, optimizer, and scheduler from checkpoint.

    Args:
        checkpoint: Checkpoint (or its path) to load.
        *args: Additional arguments to pass to `self.load`.
        **kwargs: Additional keyword arguments to pass to `self.load`.

    Raises:
        ValueError: If `model` is not defined.
        ValueError: If `model` is not an instance of `deepspeed.DeepSpeedEngine`.

    See Also:
        [`from_checkpoint`][danling.BaseRunner.from_checkpoint]: Build runner from checkpoint.
        [`load_pretrained`][danling.BaseRunner.load_pretrained]: Load model parameters from pretrained checkpoint.
    """

    if self.model is None:
        raise ValueError("model is not defined")
    if not isinstance(self.model, deepspeed.DeepSpeedEngine):
        raise ValueError("model is not an instance of `deepspeed.DeepSpeedEngine`")

    self.model.load_checkpoint(checkpoint)
    self.config.checkpoint = checkpoint

load_pretrained

Python
load_pretrained(
    checkpoint: bytes | str | PathLike, *args, **kwargs
) -> None

Load model from pretrained checkpoint.

This method only loads the model weights.

Parameters:

Name Type Description Default
checkpoint
bytes | str | PathLike

Pretrained checkpoint directory.

 required
*args

Additional arguments to pass to self.load.

 ()
**kwargs

Additional keyword arguments to pass to self.load.

 {}

Raises:

Type Description
ValueError

If model is not defined.
See Also

load_checkpoint: Load model, optimizer, and scheduler from checkpoint.

Source code in danling/runner/deepspeed_runner.py
Python
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
def load_pretrained(self, checkpoint: bytes | str | os.PathLike, *args, **kwargs) -> None:  # type: ignore[override]
    """
    Load model from pretrained checkpoint.

    This method only loads the model weights.

    Args:
        checkpoint: Pretrained checkpoint directory.
        *args: Additional arguments to pass to `self.load`.
        **kwargs: Additional keyword arguments to pass to `self.load`.

    Raises:
        ValueError: If `model` is not defined.

    See Also:
        [`load_checkpoint`][danling.BaseRunner.load_checkpoint]: Load model, optimizer, and scheduler from
            checkpoint.
    """

    if self.model is None:
        raise ValueError("model is not defined")

    self.model.load_checkpoint(checkpoint, load_module_only=True)
    self.config.pretrained = checkpoint

load_config

Python
load_config(
    checkpoint: bytes | str | PathLike,
    overwrite: bool = False,
    *args,
    **kwargs
) -> None

Load config from checkpoint.

Parameters:

Name Type Description Default
checkpoint
bytes | str | PathLike

Checkpoint (or its path) to load.

 required
overwrite
bool

If True, overwrite the current config with the loaded config. Defaults to False.

 False
*args

Additional arguments to pass to self.load.

 ()
**kwargs

Additional keyword arguments to pass to self.load.

 {}

Raises:

Type Description
FileNotFoundError

If checkpoint does not exists.
Source code in danling/runner/deepspeed_runner.py
Python
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
def load_config(
    self, checkpoint: bytes | str | os.PathLike, overwrite: bool = False, *args, **kwargs  # type: ignore[override]
) -> None:
    r"""
    Load config from checkpoint.

    Args:
        checkpoint: Checkpoint (or its path) to load.
        overwrite: If `True`, overwrite the current config with the loaded config.
            Defaults to `False`.
        *args: Additional arguments to pass to `self.load`.
        **kwargs: Additional keyword arguments to pass to `self.load`.

    Raises:
        FileNotFoundError: If `checkpoint` does not exists.
    """

    if isinstance(checkpoint, bytes):
        checkpoint = checkpoint.decode()

    config = self.load(os.path.join(checkpoint, "runner.yaml"), *args, **kwargs)
    self.config.merge(config, overwrite=overwrite)
    self.step_begin = config["steps"] + 1
    self.epoch_begin = config["epochs"] + 1

Runner

Bases: BaseRunner

Dynamic runner class that selects the appropriate platform based on configuration.

This runner dynamically modifies the __class__ attribute to adapt to the platform.

It’s safe (and recommended) to inherit from this class to extend the Runner.

Valid platform options are:

  • “auto” (default)
  • “torch”
  • “accelerate”
  • “deepspeed”

Examples:

Python Console Session
1
2
>>> config = Config({"platform": "accelerate"})
>>> runner = Runner(config)
See Also
Source code in danling/runner/runner.py
Python
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
class Runner(BaseRunner):
    r"""
    Dynamic runner class that selects the appropriate platform based on configuration.

    This runner dynamically modifies the `__class__` attribute to adapt to the platform.

    It's safe (and recommended) to inherit from this class to extend the Runner.

    Valid platform options are:

    - "auto" (default)
    - "torch"
    - "accelerate"
    - "deepspeed"

    Examples:
        >>> config = Config({"platform": "accelerate"})
        >>> runner = Runner(config)

    See Also:
        - [`BaseRunner`][danling.runner.BaseRunner]: Base class for all runners.
        - [`TorchRunner`][danling.runner.TorchRunner]: PyTorch runner.
        - [`AccelerateRunner`][danling.runner.AccelerateRunner]: PyTorch runner with Accelerate.
        - [`DeepSpeedRunner`][danling.runner.DeepSpeedRunner]: PyTorch runner with DeepSpeed.
    """

    def __init__(self, config: Config) -> None:
        platform = config.get("platform", "auto").lower()

        if platform == "auto":
            platform = "deepspeed" if ds.is_successful() else "torch"

        if platform == "accelerate":
            ac.check()
            self.__class__ = type("AccelerateRunner", (self.__class__, AccelerateRunner), {})
        elif platform == "deepspeed":
            ds.check()
            self.__class__ = type("DeepSpeedRunner", (self.__class__, DeepSpeedRunner), {})
        elif platform == "torch":
            self.__class__ = type("TorchRunner", (self.__class__, TorchRunner), {})
        else:
            raise ValueError(f"Unknown platform: {platform}. Valid options are: torch, accelerate, deepspeed")

        super().__init__(config)

TorchRunner

Bases: BaseRunner

Set up everything for running a job with PyTorch.

PyTorch backend is the most basic and flexible distributed backend. If you wish to extend the Runner, this is the best choice.

Source code in danling/runner/torch_runner.py
Python
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
class TorchRunner(BaseRunner):
    r"""
    Set up everything for running a job with PyTorch.

    PyTorch backend is the most basic and flexible distributed backend.
    If you wish to extend the Runner, this is the best choice.
    """

    model: nn.Module
    ema: nn.Module | None = None
    criterion: nn.Module
    optimizer: optim.Optimizer
    scheduler: optim.lr_scheduler._LRScheduler

    def __post_init__(self):
        super().__post_init__()
        if self.datasets:
            self.build_dataloaders()
        if self.config.get("log_interval") is None:
            self.config.log_interval = max(ceil(max(len(d) for d in self.dataloaders.values()) / 10), 1)
        self.model = self.model.to(self.device)
        if self.ema is not None:
            self.ema = self.ema.to(self.device)
        if self.distributed and not isinstance(
            self.model, (nn.parallel.DistributedDataParallel, nn.parallel.DataParallel)
        ):
            self.model = nn.parallel.DistributedDataParallel(self.model)

    def train(self, train_splits: list[str] | None = None, evaluate_splits: list[str] | None = None) -> NestedDict:
        r"""
        Perform training on `split`.

        Args:
            train_splits: list of split to run train.
                Defaults to `["train"]`.
            evaluate_splits: list of split to run evaluate.
                Defaults to `self.dataloaders` except for those in `train_splits`.

        Return:
            NestedDict: train results
        """

        early_stop_counter = 0
        if train_splits is None:
            train_splits = ["train"] if "train" in self.dataloaders else []
        self.train_splits = sorted(train_splits)
        if not train_splits:
            warn("No training split is found. Will only evaluate for one epoch.", stacklevel=2)
            self.epoch_end = self.epoch_begin + 1
        if evaluate_splits is None:
            evaluate_splits = [s for s in self.dataloaders if s not in train_splits]
        self.evaluate_splits = sorted(evaluate_splits)
        print(f"Begin training from {self.epoch_begin} to {self.epoch_end}")
        print(f"Training splits: {self.train_splits}")
        print(f"Evaluation splits: {self.evaluate_splits}")
        patience = self.config.get("patience", float("inf"))
        for epoch in range(self.epoch_begin, self.epoch_end):  # type: ignore
            self.epochs = epoch
            result = NestedDict()
            result.setattr("convert_mapping", True)
            for split in self.train_splits:
                result[split] = self.train_epoch(split)
            for split in self.evaluate_splits:
                result[split] = self.evaluate_epoch(split)
            self.append_result(result)
            print(self.format_epoch_result(result))
            self.save_result()
            if self.config.save_interval is not None:
                self.save_checkpoint()
            """@nni.report_intermediate_result(self.latest_score)"""
            early_stop_counter = 0 if self.is_best else early_stop_counter + 1
            if early_stop_counter > patience:
                print("early stop")
                break
        """@nni.report_final_result(self.latest_score)"""
        return self.results

    def train_epoch(self, split: str = "train") -> NestedDict:
        r"""
        Train one epoch on `split`.

        Args:
            split (str): split to run train

        Return:
            NestedDict: train result
        """

        self.mode = "train"  # type: ignore
        self.split = split
        loader = self.dataloaders[split]
        length = len(loader) - 1
        last_print_iteration = -1
        log_interval = self.config.get("log_interval", -1)
        self.meters.reset()
        if self.train_metrics is not None:
            self.metrics = self.train_metrics
        if self.metrics is not None:
            self.metrics.reset()
        batch_time = time()
        if hasattr(loader.batch_sampler, "set_epoch"):
            loader.batch_sampler.set_epoch(self.epochs)
        if hasattr(loader.sampler, "set_epoch"):
            loader.sampler.set_epoch(self.epochs)

        for iteration, data in enumerate(loader):
            _, loss = self.train_step(data)

            if log_interval > 0 and (iteration > 0 and iteration % log_interval == 0 or iteration == length):
                interval = iteration - last_print_iteration
                if self.device == torch.device("cuda"):
                    torch.cuda.synchronize()
                if self.scheduler is not None:
                    self.meters.lr.update(self.scheduler.get_last_lr()[0])
                self.meters.time.update((time() - batch_time) / interval)
                batch_time = time()
                reduced_loss = self.reduce(loss).item()
                self.meters.loss.update(reduced_loss)
                self.step_log(split, iteration, length)
                last_print_iteration = iteration

        result = self.get_epoch_result()
        return result

    def train_step(self, data) -> Tuple[Any, torch.Tensor]:
        with self.autocast(), self.accumulate():
            input = data["input"] if isinstance(data, Mapping) else data[0]
            target = data["target"] if isinstance(data, Mapping) else data[1]
            pred = self.model(**input) if isinstance(input, Mapping) else self.model(input)
            loss = self.criterion(pred, target)
            if self.metrics is not None:
                self.metrics.update(pred.squeeze(-1), target)
            self.advance(loss)
        return pred, loss

    def advance(self, loss) -> None:
        r"""
        Backward loss and step optimizer & scheduler.

        Args:
            loss: The loss tensor from which to backpropagate.
        """

        self.backward(loss / self.accum_steps)
        if self.accum_steps <= 1 or self.steps % self.accum_steps == 0:
            if self.config.get("max_grad_value") is not None:
                clip_grad_value_(self.model.parameters(), self.config["max_grad_value"])
            if self.config.get("max_grad_norm") is not None:
                clip_grad_norm_(self.model.parameters(), self.config["max_grad_norm"])
            self.optimizer.step()
            self.optimizer.zero_grad()
            if self.ema is not None:
                self.ema.update()
            if self.scheduler is not None:
                self.scheduler.step()
            self.steps += 1

    def evaluate(self, evaluate_splits: list[str] | None = None) -> NestedDict:
        r"""
        Perform evaluation on `evaluate_splits`.

        Args:
            evaluate_splits: list of split to run evaluate.
                Defaults to `["evaluate"]`.

        Return:
            NestedDict: evaluation result
        """

        if evaluate_splits is None:
            evaluate_splits = ["evaluate"]
        self.evaluate_splits = sorted(evaluate_splits)

        print("Begin evaluation")
        print(f"Evaluation splits: {self.evaluate_splits}")
        result = NestedDict()
        result.setattr("convert_mapping", True)
        for split in self.evaluate_splits:
            result[split] = self.evaluate_epoch(split=split)
        print(self.format_epoch_result(result))
        return result

    # torch.inference_mode cause experiments to hang
    # @torch.inference_mode()
    def evaluate_epoch(self, split: str = "val") -> NestedDict:
        r"""
        Evaluate one epoch on `split`.

        Args:
            split (str): split to run evaluate

        Return:
            NestedDict: evaluation result
        """

        self.mode = RunnerMode.evaluate
        self.split = split
        loader = self.dataloaders[split]
        length = len(loader) - 1
        last_print_iteration = -1
        log_interval = self.config.get("log_interval", -1)
        self.meters.reset()
        if self.evaluate_metrics is not None:
            self.metrics = self.evaluate_metrics
        if self.metrics is not None:
            self.metrics.reset()
        batch_time = time()

        for iteration, data in enumerate(loader):
            _, loss = self.evaluate_step(data)

            if log_interval > 0 and (iteration > 0 and iteration % log_interval == 0 or iteration == length):
                interval = iteration - last_print_iteration
                if self.device == torch.device("cuda"):
                    torch.cuda.synchronize()
                self.meters.time.update((time() - batch_time) / interval)
                batch_time = time()
                reduced_loss = self.reduce(loss).item()
                self.meters.loss.update(reduced_loss)
                self.step_log(split, iteration, length)
                last_print_iteration = iteration

        result = self.get_epoch_result()
        self.write_result(result, split, self.epochs)
        return result

    def evaluate_step(self, data) -> Tuple[Any, torch.Tensor]:
        input = data["input"] if isinstance(data, Mapping) else data[0]
        target = data["target"] if isinstance(data, Mapping) else data[1]
        model = self.ema or self.model
        pred = model(**input) if isinstance(input, Mapping) else model(input)
        loss = self.criterion(pred, target)
        if self.metrics is not None:
            self.metrics.update(pred.squeeze(-1), target)
        return pred, loss

    @torch.inference_mode()
    def infer(self, split: str = "infer") -> list[float]:
        r"""
        Perform inference on `split`.

        Args:
            split (str): split to run inference

        Return:
            Tensor: inference outputs
        """

        self.mode = RunnerMode.infer
        loader = self.dataloaders[split]
        output: list[float] = []
        model = self.ema or self.model
        for _, data in tqdm(enumerate(loader), total=len(loader)):
            input = data["input"] if isinstance(data, Mapping) else data[0]
            pred = model(**input) if isinstance(input, Mapping) else model(input)
            output.extend(pred.squeeze(-1).tolist())

        if self.distributed:
            torch.cuda.synchronize()
            output = self.gather_for_metrics(output)
        return output

    def backward(self, loss: torch.Tensor) -> None:
        r"""
        Backward loss.

        Args:
            loss: Loss to backward.
        """

        loss.backward()

    def has_nan_inf_grad(self, model: nn.Module | None = None) -> bool:
        r"""
        Check if model has NaN or Inf gradients.

        Args:
            model: Model to check.
                Defaults to `self.model`.

        Return:
            bool: True if NaN or Inf is detected in gradients.
        """
        model = model or self.model
        for name, param in model.named_parameters():
            if param.grad is not None:
                if torch.isnan(param.grad).any():
                    print(f"NaN detected in gradients of parameter: {name}")
                    return True
                if torch.isinf(param.grad).any():
                    print(f"Inf detected in gradients of parameter: {name}")
                    return True
        return False

    def init_distributed(self) -> None:
        r"""
        Set up distributed training.

        Initialise process group and set up DDP variables.
        """

        backend = self.config.get("backend", os.getenv("BACKEND"))
        init_method = self.config.get("init_method", os.getenv("INIT_METHOD"))
        world_size = int(self.config.get("world_size", os.getenv("WORLD_SIZE", "1")))
        rank = int(self.config.get("rank", os.getenv("RANK", "0")))
        if world_size > 1:
            if torch.cuda.is_available():
                torch.cuda.set_device(self.local_rank)
            dist.init_process_group(backend, init_method, world_size=world_size, rank=rank)
            object_list = [self.id, self.timestamp]
            dist.broadcast_object_list(object_list)
            self.id, self.timestamp = object_list

    @on_main_process
    def init_tensorboard(self, *args, **kwargs) -> None:
        r"""
        Set up Tensoraoard SummaryWriter.
        """
        from torch.utils.tensorboard.writer import SummaryWriter  # pylint: disable=C0415

        if "log_dir" not in kwargs:
            kwargs["log_dir"] = self.dir

        self.writer = SummaryWriter(*args, **kwargs)
        self.writer.add_scalar = catch(OSError, verbose=False)(self.writer.add_scalar)

    def set_seed(self, seed: int = None, bias: int = None) -> int:  # type: ignore[assignment]
        r"""
        Set up random seed.

        Args:
            seed: Random seed to set.
                Defaults to `self.config.seed` (`config.seed`).

            bias: Make the seed different for each processes.
                This is used to ensure the data augmentation are applied differently on every processes.
                Defaults to `self.rank`.
                Set to `False` to disable this feature.
        Returns:
            Random seed set.
        """

        seed = seed or self.config.seed  # type: ignore[assignment]
        if seed is None:
            if self.inited:
                seed = random.randint(0, 2**32 - 1)
                if self.distributed:
                    object_list = [seed]
                    dist.broadcast_object_list(object_list)
                    seed = object_list[0]
                self.config.seed = seed
        else:
            seed = defaults.SEED
        bias = bias or self.rank
        if bias:
            seed += bias
        torch.manual_seed(seed)
        torch.cuda.manual_seed(seed)
        if np_random is not None:
            np_random.seed(seed)
        random.seed(seed)
        return seed

    def set_deterministic(self) -> None:
        cudnn.benchmark = False
        cudnn.deterministic = True
        if torch.__version__ >= "1.8.0":
            torch.use_deterministic_algorithms(True)

    def state_dict(self, cls: Callable = dict) -> Mapping:
        if self.model is None:
            raise ValueError("Model must be defined when calling state_dict")
        return cls(
            runner=self.config.dict(),
            model=self.unwrap(self.model).state_dict(),
            optimizer=self.optimizer.state_dict() if self.optimizer else None,
            scheduler=self.scheduler.state_dict() if self.scheduler else None,
        )

    def unwrap(self, model: nn.Module) -> nn.Module:
        if isinstance(model, (nn.parallel.DistributedDataParallel, nn.parallel.DataParallel)):
            return model.module
        return model

    def build_dataloaders(self):
        datasets = {k: d for k, d in self.datasets.items() if k not in self.dataloaders}
        default_kwargs = self.config.get("dataloader", NestedDict())
        dataloader_kwargs = NestedDict({k: default_kwargs.pop(k) for k in self.datasets if k in default_kwargs})
        for k, d in datasets.items():
            dataloader_kwargs.setdefault(k, NestedDict())
            dataloader_kwargs[k].merge(default_kwargs, overwrite=False)
            shuffle = dataloader_kwargs[k].pop("shuffle", getattr(d, "train", True))
            if self.distributed:
                sampler = utils.data.distributed.DistributedSampler(d, shuffle=shuffle)
            else:
                sampler = utils.data.RandomSampler(d) if shuffle else utils.data.SequentialSampler(d)
            dataloader_kwargs[k].setdefault("drop_last", not getattr(d, "train", True))
            self.dataloaders[k] = utils.data.DataLoader(
                d, sampler=sampler, collate_fn=self.collate_fn, **dataloader_kwargs[k]
            )

    def collate_fn(self, batch):
        return to_device(utils.data.dataloader.default_collate(batch), self.device)

    @contextmanager
    def autocast(self):
        if self.config.get("precision") is None:
            yield nullcontext()
        else:
            yield torch.autocast(self.device.type, dtype=get_precision(self.config.precision))

    @contextmanager
    def accumulate(self):
        if self.accum_steps <= 1 or self.steps % self.accum_steps == 0:
            yield nullcontext()
        else:
            yield self.model.no_sync()

    def get_optimizer(self, name: str):
        if name.lower() == "sgd":
            return optim.SGD
        if name.lower() == "asgd":
            return optim.ASGD
        if name.lower() in {"torch_adam", "torch_adamw"}:
            return optim.Adam
        if ds is not None:
            if name.lower() == "adagrad":
                return ds.ops.adagrad.DeepSpeedCPUAdagrad
            if name.lower() in {"adam", "adamw"}:
                if torch.cuda.device_count() > 0:
                    return ds.ops.adam.FusedAdam
                return ds.ops.adam.DeepSpeedCPUAdam
            if name.lower() in {"cpu", "cpu_adam", "cpuadam", "cpu_adamw", "cpuadamw"}:
                return ds.ops.adam.DeepSpeedCPUAdam
            if name.lower() == "lamb":
                if torch.cuda.device_count() > 0:
                    return ds.ops.lamb.FusedLamb
                return ds.ops.lamb.DeepSpeedCPULamb
            if name.lower() in {"cpulamb", "cpu_lamb"}:
                return ds.ops.lamb.DeepSpeedCPULamb
            if name.lower() == "lion":
                if torch.cuda.device_count() > 0:
                    return ds.ops.lion.FusedLion
                return ds.ops.lion.DeepSpeedCPULion
            if name.lower() in {"cpulion", "cpu_lion"}:
                return ds.ops.lion.DeepSpeedCPULion
        if name.lower() in {"adam", "adamw"}:
            return optim.AdamW
        if name.lower() == "adadelta":
            return optim.Adadelta
        if name.lower() == "adafactor":
            return optim.Adafactor
        if name.lower() == "adagrad":
            return optim.Adagrad
        if name.lower() == "adamax":
            return optim.Adamax
        if name.lower() == "lbfgs":
            return optim.LBFGS
        if name.lower() == "nadam":
            return optim.NAdam
        if name.lower() == "radam":
            return optim.RAdam
        if name.lower() == "rmsprop":
            return optim.RMSprop
        if name.lower() == "rprop":
            return optim.Rprop

    def get_deepspeed_config(self, config: NestedDict | str = None) -> NestedDict:  # pylint: disable=R0912, R0915
        r"""
        Preprocess DeepSpeed config.
        """

        if config is None and "deepspeed" in self.config:
            config = self.config.deepspeed
        if isinstance(config, str):
            config = NestedDict(config)
        if config is None:
            config = NestedDict()
        if config.get("steps_per_print", "auto") == "auto":
            config["steps_per_print"] = self.config.log_interval
        if config.get("train_micro_batch_size_per_gpu", "auto") == "auto":
            config["train_micro_batch_size_per_gpu"] = self.batch_size
        if config.get("gradient_accumulation_steps", "auto") == "auto":
            if self.accum_steps > 1:
                config["gradient_accumulation_steps"] = self.accum_steps
            else:
                config.pop("gradient_accumulation_steps", None)
        if "amp" in config:
            amp = config["amp"]
            if amp.get("enabled", "auto") == "auto":
                amp["enabled"] = "true"
            if amp.get("opt_level", "auto") == "auto":
                amp["opt_level"] = "O1"
        if "zero_optimization" in config:
            zero = config["zero_optimization"]
            if zero.get("allgather_bucket_size") == "auto":
                zero["allgather_bucket_size"] = 1e6
            if zero.get("reduce_bucket_size") == "auto":
                zero["reduce_bucket_size"] = 1e6
            if zero.get("stage3_max_live_parameters") == "auto":
                zero["stage3_max_live_parameters"] = 1e8
            if zero.get("stage3_max_live_gradients") == "auto":
                zero["stage3_max_live_gradients"] = 1e8
            if zero.get("stage3_max_reuse_distance") == "auto":
                zero["stage3_max_reuse_distance"] = 1e8
            if zero.get("stage3_prefetch_bucket_size") == "auto":
                zero["stage3_prefetch_bucket_size"] = 1e6
            if zero.get("stage3_param_persistence_threshold") == "auto":
                zero["stage3_param_persistence_threshold"] = 1e8
            if "amp" in config:
                if "fp16" not in config:
                    config["fp16"] = NestedDict()
                if config["fp16"].get("enabled", "auto"):
                    config["fp16"]["enabled"] = config["amp"]["enabled"]
                warn(
                    f"AMP is not compatible with ZeRO. Automatically set 'fp16' to {config['amp']['enabled']}",
                    stacklevel=2,
                )
                del config["amp"]
        if "optimizer" in config:
            if config["optimizer"].get("type", "auto") == "auto":
                config["optimizer"]["type"] = "Adam"
            if "params" not in config["optimizer"]:
                config["optimizer"]["params"] = NestedDict()
            optimizer = config["optimizer"]["params"]
            if optimizer.get("lr", "auto") == "auto":
                optimizer["lr"] = self.config.get("optim.lr", 1e-3)
            if optimizer.get("weight_decay", "auto") == "auto":
                optimizer["weight_decay"] = self.config.get("optim.weight_decay", 1e-2)
            if optimizer.get("betas") == "auto":
                optimizer["betas"] = (0.9, 0.999)
            if optimizer.get("eps") == "auto":
                optimizer["eps"] = 1e-8
        if "scheduler" in config:
            if config["scheduler"].get("type", "auto") == "auto":
                config["scheduler"]["type"] = "WarmupCosineLR"
            if "params" not in config["scheduler"]:
                config["scheduler"]["params"] = NestedDict()
            scheduler = config["scheduler"]["params"]
            if scheduler.get("total_num_steps", "auto") == "auto":
                scheduler["total_num_steps"] = self.total_steps
            if scheduler.get("warmup_num_steps", "auto") == "auto":
                scheduler["warmup_num_steps"] = scheduler["total_num_steps"] // 20
            if config["scheduler"]["type"] in ("WarmupLR", "WarmupDecayLR"):
                if scheduler.get("warmup_max_lr", "auto") == "auto":
                    if self.optimizer:
                        scheduler["warmup_max_lr"] = self.optimizer.param_groups[0]["lr"]
                    elif "optimizer" in config:
                        scheduler["warmup_max_lr"] = config["optimizer"]["params"]["lr"]
                    else:
                        scheduler["warmup_max_lr"] = self.config.get("optim.lr", 1e-3)
                if scheduler.get("warmup_min_lr", "auto") == "auto":
                    scheduler["warmup_min_lr"] = 1e-9
            else:
                scheduler.pop("warmup_max_lr", None)
                scheduler.pop("warmup_min_lr", None)
        if config.get("gradient_clipping", "auto") == "auto" and self.config.get("max_grad_norm") is not None:
            config["gradient_clipping"] = self.config["max_grad_norm"]
        return config

    @property
    def device(self):
        return torch.device("cuda" if torch.cuda.is_available() else "cpu", self.local_rank)

    @property
    def mode(self) -> RunnerMode:
        return self._mode

    @mode.setter
    def mode(self, mode: str | RunnerMode) -> None:
        if isinstance(mode, str):
            mode = RunnerMode(mode)
        self._mode = mode
        if self.model is not None:
            self.model.train(mode == RunnerMode.train)
        if self.ema is not None:
            self.ema.train(mode == RunnerMode.train)

    @property
    def rank(self) -> int:
        if self.distributed:
            return dist.get_rank()
        return 0

    @property
    def local_rank(self) -> int:
        if local_rank := os.getenv("LOCAL_RANK"):
            return int(local_rank)
        return 0

    @property
    def world_size(self) -> int:
        r"""
        Number of Processes.
        """

        if torch.distributed.is_available() and torch.distributed.is_initialized():
            return dist.get_world_size()
        return 1

    @property
    def distributed(self) -> bool:
        return self.world_size > 1

    @cached_property
    def accum_steps(self) -> int:
        return self.config.get("accum_steps", 1)

    @staticmethod
    def reduce(tensor: torch.Tensor) -> torch.Tensor:
        if torch.distributed.is_available() and torch.distributed.is_initialized():
            dist.all_reduce(tensor)
        return tensor

world_size property

Python
world_size: int

Number of Processes.

train

Python
train(
    train_splits: list[str] | None = None,
    evaluate_splits: list[str] | None = None,
) -> NestedDict

Perform training on split.

Parameters:

Name Type Description Default
train_splits
list[str] | None

list of split to run train. Defaults to ["train"].

 None
evaluate_splits
list[str] | None

list of split to run evaluate. Defaults to self.dataloaders except for those in train_splits.

 None
Return
Source code in danling/runner/torch_runner.py
Python
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
def train(self, train_splits: list[str] | None = None, evaluate_splits: list[str] | None = None) -> NestedDict:
    r"""
    Perform training on `split`.

    Args:
        train_splits: list of split to run train.
            Defaults to `["train"]`.
        evaluate_splits: list of split to run evaluate.
            Defaults to `self.dataloaders` except for those in `train_splits`.

    Return:
        NestedDict: train results
    """

    early_stop_counter = 0
    if train_splits is None:
        train_splits = ["train"] if "train" in self.dataloaders else []
    self.train_splits = sorted(train_splits)
    if not train_splits:
        warn("No training split is found. Will only evaluate for one epoch.", stacklevel=2)
        self.epoch_end = self.epoch_begin + 1
    if evaluate_splits is None:
        evaluate_splits = [s for s in self.dataloaders if s not in train_splits]
    self.evaluate_splits = sorted(evaluate_splits)
    print(f"Begin training from {self.epoch_begin} to {self.epoch_end}")
    print(f"Training splits: {self.train_splits}")
    print(f"Evaluation splits: {self.evaluate_splits}")
    patience = self.config.get("patience", float("inf"))
    for epoch in range(self.epoch_begin, self.epoch_end):  # type: ignore
        self.epochs = epoch
        result = NestedDict()
        result.setattr("convert_mapping", True)
        for split in self.train_splits:
            result[split] = self.train_epoch(split)
        for split in self.evaluate_splits:
            result[split] = self.evaluate_epoch(split)
        self.append_result(result)
        print(self.format_epoch_result(result))
        self.save_result()
        if self.config.save_interval is not None:
            self.save_checkpoint()
        """@nni.report_intermediate_result(self.latest_score)"""
        early_stop_counter = 0 if self.is_best else early_stop_counter + 1
        if early_stop_counter > patience:
            print("early stop")
            break
    """@nni.report_final_result(self.latest_score)"""
    return self.results

train_epoch

Python
train_epoch(split: str = 'train') -> NestedDict

Train one epoch on split.

Parameters:

Name Type Description Default
split
str

split to run train

 'train'
Return
Source code in danling/runner/torch_runner.py
Python
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
def train_epoch(self, split: str = "train") -> NestedDict:
    r"""
    Train one epoch on `split`.

    Args:
        split (str): split to run train

    Return:
        NestedDict: train result
    """

    self.mode = "train"  # type: ignore
    self.split = split
    loader = self.dataloaders[split]
    length = len(loader) - 1
    last_print_iteration = -1
    log_interval = self.config.get("log_interval", -1)
    self.meters.reset()
    if self.train_metrics is not None:
        self.metrics = self.train_metrics
    if self.metrics is not None:
        self.metrics.reset()
    batch_time = time()
    if hasattr(loader.batch_sampler, "set_epoch"):
        loader.batch_sampler.set_epoch(self.epochs)
    if hasattr(loader.sampler, "set_epoch"):
        loader.sampler.set_epoch(self.epochs)

    for iteration, data in enumerate(loader):
        _, loss = self.train_step(data)

        if log_interval > 0 and (iteration > 0 and iteration % log_interval == 0 or iteration == length):
            interval = iteration - last_print_iteration
            if self.device == torch.device("cuda"):
                torch.cuda.synchronize()
            if self.scheduler is not None:
                self.meters.lr.update(self.scheduler.get_last_lr()[0])
            self.meters.time.update((time() - batch_time) / interval)
            batch_time = time()
            reduced_loss = self.reduce(loss).item()
            self.meters.loss.update(reduced_loss)
            self.step_log(split, iteration, length)
            last_print_iteration = iteration

    result = self.get_epoch_result()
    return result

advance

Python
advance(loss) -> None

Backward loss and step optimizer & scheduler.

Parameters:

Name Type Description Default
loss

The loss tensor from which to backpropagate.

 required
Source code in danling/runner/torch_runner.py
Python
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
def advance(self, loss) -> None:
    r"""
    Backward loss and step optimizer & scheduler.

    Args:
        loss: The loss tensor from which to backpropagate.
    """

    self.backward(loss / self.accum_steps)
    if self.accum_steps <= 1 or self.steps % self.accum_steps == 0:
        if self.config.get("max_grad_value") is not None:
            clip_grad_value_(self.model.parameters(), self.config["max_grad_value"])
        if self.config.get("max_grad_norm") is not None:
            clip_grad_norm_(self.model.parameters(), self.config["max_grad_norm"])
        self.optimizer.step()
        self.optimizer.zero_grad()
        if self.ema is not None:
            self.ema.update()
        if self.scheduler is not None:
            self.scheduler.step()
        self.steps += 1

evaluate

Python
evaluate(
    evaluate_splits: list[str] | None = None,
) -> NestedDict

Perform evaluation on evaluate_splits.

Parameters:

Name Type Description Default
evaluate_splits
list[str] | None

list of split to run evaluate. Defaults to ["evaluate"].

 None
Return
Source code in danling/runner/torch_runner.py
Python
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
def evaluate(self, evaluate_splits: list[str] | None = None) -> NestedDict:
    r"""
    Perform evaluation on `evaluate_splits`.

    Args:
        evaluate_splits: list of split to run evaluate.
            Defaults to `["evaluate"]`.

    Return:
        NestedDict: evaluation result
    """

    if evaluate_splits is None:
        evaluate_splits = ["evaluate"]
    self.evaluate_splits = sorted(evaluate_splits)

    print("Begin evaluation")
    print(f"Evaluation splits: {self.evaluate_splits}")
    result = NestedDict()
    result.setattr("convert_mapping", True)
    for split in self.evaluate_splits:
        result[split] = self.evaluate_epoch(split=split)
    print(self.format_epoch_result(result))
    return result

evaluate_epoch

Python
evaluate_epoch(split: str = 'val') -> NestedDict

Evaluate one epoch on split.

Parameters:

Name Type Description Default
split
str

split to run evaluate

 'val'
Return
Source code in danling/runner/torch_runner.py
Python
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
def evaluate_epoch(self, split: str = "val") -> NestedDict:
    r"""
    Evaluate one epoch on `split`.

    Args:
        split (str): split to run evaluate

    Return:
        NestedDict: evaluation result
    """

    self.mode = RunnerMode.evaluate
    self.split = split
    loader = self.dataloaders[split]
    length = len(loader) - 1
    last_print_iteration = -1
    log_interval = self.config.get("log_interval", -1)
    self.meters.reset()
    if self.evaluate_metrics is not None:
        self.metrics = self.evaluate_metrics
    if self.metrics is not None:
        self.metrics.reset()
    batch_time = time()

    for iteration, data in enumerate(loader):
        _, loss = self.evaluate_step(data)

        if log_interval > 0 and (iteration > 0 and iteration % log_interval == 0 or iteration == length):
            interval = iteration - last_print_iteration
            if self.device == torch.device("cuda"):
                torch.cuda.synchronize()
            self.meters.time.update((time() - batch_time) / interval)
            batch_time = time()
            reduced_loss = self.reduce(loss).item()
            self.meters.loss.update(reduced_loss)
            self.step_log(split, iteration, length)
            last_print_iteration = iteration

    result = self.get_epoch_result()
    self.write_result(result, split, self.epochs)
    return result

infer

Python
infer(split: str = 'infer') -> list[float]

Perform inference on split.

Parameters:

Name Type Description Default
split
str

split to run inference

 'infer'
Return
Source code in danling/runner/torch_runner.py
Python
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
@torch.inference_mode()
def infer(self, split: str = "infer") -> list[float]:
    r"""
    Perform inference on `split`.

    Args:
        split (str): split to run inference

    Return:
        Tensor: inference outputs
    """

    self.mode = RunnerMode.infer
    loader = self.dataloaders[split]
    output: list[float] = []
    model = self.ema or self.model
    for _, data in tqdm(enumerate(loader), total=len(loader)):
        input = data["input"] if isinstance(data, Mapping) else data[0]
        pred = model(**input) if isinstance(input, Mapping) else model(input)
        output.extend(pred.squeeze(-1).tolist())

    if self.distributed:
        torch.cuda.synchronize()
        output = self.gather_for_metrics(output)
    return output

backward

Python
backward(loss: Tensor) -> None

Backward loss.

Parameters:

Name Type Description Default
loss
Tensor

Loss to backward.

 required
Source code in danling/runner/torch_runner.py
Python
324
325
326
327
328
329
330
331
332
def backward(self, loss: torch.Tensor) -> None:
    r"""
    Backward loss.

    Args:
        loss: Loss to backward.
    """

    loss.backward()

has_nan_inf_grad

Python
has_nan_inf_grad(model: Module | None = None) -> bool

Check if model has NaN or Inf gradients.

Parameters:

Name Type Description Default
model
Module | None

Model to check. Defaults to self.model.

 None
Return
Source code in danling/runner/torch_runner.py
Python
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
def has_nan_inf_grad(self, model: nn.Module | None = None) -> bool:
    r"""
    Check if model has NaN or Inf gradients.

    Args:
        model: Model to check.
            Defaults to `self.model`.

    Return:
        bool: True if NaN or Inf is detected in gradients.
    """
    model = model or self.model
    for name, param in model.named_parameters():
        if param.grad is not None:
            if torch.isnan(param.grad).any():
                print(f"NaN detected in gradients of parameter: {name}")
                return True
            if torch.isinf(param.grad).any():
                print(f"Inf detected in gradients of parameter: {name}")
                return True
    return False

init_distributed

Python
init_distributed() -> None

Set up distributed training.

Initialise process group and set up DDP variables.

Source code in danling/runner/torch_runner.py
Python
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
def init_distributed(self) -> None:
    r"""
    Set up distributed training.

    Initialise process group and set up DDP variables.
    """

    backend = self.config.get("backend", os.getenv("BACKEND"))
    init_method = self.config.get("init_method", os.getenv("INIT_METHOD"))
    world_size = int(self.config.get("world_size", os.getenv("WORLD_SIZE", "1")))
    rank = int(self.config.get("rank", os.getenv("RANK", "0")))
    if world_size > 1:
        if torch.cuda.is_available():
            torch.cuda.set_device(self.local_rank)
        dist.init_process_group(backend, init_method, world_size=world_size, rank=rank)
        object_list = [self.id, self.timestamp]
        dist.broadcast_object_list(object_list)
        self.id, self.timestamp = object_list

init_tensorboard

Python
init_tensorboard(*args, **kwargs) -> None

Set up Tensoraoard SummaryWriter.

Source code in danling/runner/torch_runner.py
Python
375
376
377
378
379
380
381
382
383
384
385
386
@on_main_process
def init_tensorboard(self, *args, **kwargs) -> None:
    r"""
    Set up Tensoraoard SummaryWriter.
    """
    from torch.utils.tensorboard.writer import SummaryWriter  # pylint: disable=C0415

    if "log_dir" not in kwargs:
        kwargs["log_dir"] = self.dir

    self.writer = SummaryWriter(*args, **kwargs)
    self.writer.add_scalar = catch(OSError, verbose=False)(self.writer.add_scalar)

set_seed

Python
set_seed(seed: int = None, bias: int = None) -> int

Set up random seed.

Parameters:

Name Type Description Default
seed
int

Random seed to set. Defaults to self.config.seed (config.seed).

 None
bias
int

Make the seed different for each processes. This is used to ensure the data augmentation are applied differently on every processes. Defaults to self.rank. Set to False to disable this feature.

 None
Source code in danling/runner/torch_runner.py
Python
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
def set_seed(self, seed: int = None, bias: int = None) -> int:  # type: ignore[assignment]
    r"""
    Set up random seed.

    Args:
        seed: Random seed to set.
            Defaults to `self.config.seed` (`config.seed`).

        bias: Make the seed different for each processes.
            This is used to ensure the data augmentation are applied differently on every processes.
            Defaults to `self.rank`.
            Set to `False` to disable this feature.
    Returns:
        Random seed set.
    """

    seed = seed or self.config.seed  # type: ignore[assignment]
    if seed is None:
        if self.inited:
            seed = random.randint(0, 2**32 - 1)
            if self.distributed:
                object_list = [seed]
                dist.broadcast_object_list(object_list)
                seed = object_list[0]
            self.config.seed = seed
    else:
        seed = defaults.SEED
    bias = bias or self.rank
    if bias:
        seed += bias
    torch.manual_seed(seed)
    torch.cuda.manual_seed(seed)
    if np_random is not None:
        np_random.seed(seed)
    random.seed(seed)
    return seed

get_deepspeed_config

Python
get_deepspeed_config(
    config: NestedDict | str = None,
) -> NestedDict

Preprocess DeepSpeed config.

Source code in danling/runner/torch_runner.py
Python
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
def get_deepspeed_config(self, config: NestedDict | str = None) -> NestedDict:  # pylint: disable=R0912, R0915
    r"""
    Preprocess DeepSpeed config.
    """

    if config is None and "deepspeed" in self.config:
        config = self.config.deepspeed
    if isinstance(config, str):
        config = NestedDict(config)
    if config is None:
        config = NestedDict()
    if config.get("steps_per_print", "auto") == "auto":
        config["steps_per_print"] = self.config.log_interval
    if config.get("train_micro_batch_size_per_gpu", "auto") == "auto":
        config["train_micro_batch_size_per_gpu"] = self.batch_size
    if config.get("gradient_accumulation_steps", "auto") == "auto":
        if self.accum_steps > 1:
            config["gradient_accumulation_steps"] = self.accum_steps
        else:
            config.pop("gradient_accumulation_steps", None)
    if "amp" in config:
        amp = config["amp"]
        if amp.get("enabled", "auto") == "auto":
            amp["enabled"] = "true"
        if amp.get("opt_level", "auto") == "auto":
            amp["opt_level"] = "O1"
    if "zero_optimization" in config:
        zero = config["zero_optimization"]
        if zero.get("allgather_bucket_size") == "auto":
            zero["allgather_bucket_size"] = 1e6
        if zero.get("reduce_bucket_size") == "auto":
            zero["reduce_bucket_size"] = 1e6
        if zero.get("stage3_max_live_parameters") == "auto":
            zero["stage3_max_live_parameters"] = 1e8
        if zero.get("stage3_max_live_gradients") == "auto":
            zero["stage3_max_live_gradients"] = 1e8
        if zero.get("stage3_max_reuse_distance") == "auto":
            zero["stage3_max_reuse_distance"] = 1e8
        if zero.get("stage3_prefetch_bucket_size") == "auto":
            zero["stage3_prefetch_bucket_size"] = 1e6
        if zero.get("stage3_param_persistence_threshold") == "auto":
            zero["stage3_param_persistence_threshold"] = 1e8
        if "amp" in config:
            if "fp16" not in config:
                config["fp16"] = NestedDict()
            if config["fp16"].get("enabled", "auto"):
                config["fp16"]["enabled"] = config["amp"]["enabled"]
            warn(
                f"AMP is not compatible with ZeRO. Automatically set 'fp16' to {config['amp']['enabled']}",
                stacklevel=2,
            )
            del config["amp"]
    if "optimizer" in config:
        if config["optimizer"].get("type", "auto") == "auto":
            config["optimizer"]["type"] = "Adam"
        if "params" not in config["optimizer"]:
            config["optimizer"]["params"] = NestedDict()
        optimizer = config["optimizer"]["params"]
        if optimizer.get("lr", "auto") == "auto":
            optimizer["lr"] = self.config.get("optim.lr", 1e-3)
        if optimizer.get("weight_decay", "auto") == "auto":
            optimizer["weight_decay"] = self.config.get("optim.weight_decay", 1e-2)
        if optimizer.get("betas") == "auto":
            optimizer["betas"] = (0.9, 0.999)
        if optimizer.get("eps") == "auto":
            optimizer["eps"] = 1e-8
    if "scheduler" in config:
        if config["scheduler"].get("type", "auto") == "auto":
            config["scheduler"]["type"] = "WarmupCosineLR"
        if "params" not in config["scheduler"]:
            config["scheduler"]["params"] = NestedDict()
        scheduler = config["scheduler"]["params"]
        if scheduler.get("total_num_steps", "auto") == "auto":
            scheduler["total_num_steps"] = self.total_steps
        if scheduler.get("warmup_num_steps", "auto") == "auto":
            scheduler["warmup_num_steps"] = scheduler["total_num_steps"] // 20
        if config["scheduler"]["type"] in ("WarmupLR", "WarmupDecayLR"):
            if scheduler.get("warmup_max_lr", "auto") == "auto":
                if self.optimizer:
                    scheduler["warmup_max_lr"] = self.optimizer.param_groups[0]["lr"]
                elif "optimizer" in config:
                    scheduler["warmup_max_lr"] = config["optimizer"]["params"]["lr"]
                else:
                    scheduler["warmup_max_lr"] = self.config.get("optim.lr", 1e-3)
            if scheduler.get("warmup_min_lr", "auto") == "auto":
                scheduler["warmup_min_lr"] = 1e-9
        else:
            scheduler.pop("warmup_max_lr", None)
            scheduler.pop("warmup_min_lr", None)
    if config.get("gradient_clipping", "auto") == "auto" and self.config.get("max_grad_norm") is not None:
        config["gradient_clipping"] = self.config["max_grad_norm"]
    return config

NestedTensor

A container for variable-length tensors that enables efficient batch operations.

NestedTensor solves a fundamental problem in deep learning: handling sequences of different lengths in batch operations. Instead of excessive padding or complex bucketing, NestedTensor provides an elegant solution that maintains both efficiency and usability.

The class provides three main views of the data: - .tensor: A padded tensor with zeros (or other value) in place of missing elements - .mask: A boolean mask indicating which elements are real vs padding - .concat: A flattened tensor containing all elements concatenated (no padding)

When indexing a NestedTensor, the behavior depends on the index type: 1. Integer index (nt[0]): Returns a single tensor without padding 2. Slice index (nt[:]): Returns a tuple of (padded_tensor, mask) 3. Tuple index (nt[:, 1:]): Returns a new NestedTensor with the specified sliced shape

Attributes:

Name Type Description
_storage

The sequence of original tensors (internal use)
tensor Tensor

Padded tensor representing all sequences with padding
mask Tensor

Boolean mask where True indicates real elements, False indicates padding
concat Tensor

Concatenated tensor of all sequences without padding
batch_first bool

Whether the first dimension is the batch dimension (B, N, *) If False, the first dimension is the sequence dimension (N, B, *)
padding_value SupportsFloat

Value used for padding in the padded tensor
mask_value bool

Value used in the mask to indicate padding positions (usually False)

Parameters:

Name Type Description Default

*tensors

Iterable[Tensor]

Variable-length tensors or sequences to store

 ()

batch_first

bool

Whether to use batch-first representation.

 True

padding_value

SupportsFloat

Value to use for padding.

 0.0

mask_value

bool

Value to use for padding positions in mask.

 False

Raises:

Type Description
ValueError

If tensors is not an iterable or is empty

Examples:

Basic usage:

Python Console Session
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
>>> nested_tensor = NestedTensor(torch.tensor([1, 2, 3]), torch.tensor([4, 5]))
>>> nested_tensor.shape
torch.Size([2, 3])
>>> nested_tensor.tensor  # Padded representation
tensor([[1, 2, 3],
        [4, 5, 0]])
>>> nested_tensor.mask  # Mask showing real vs padding values
tensor([[ True,  True,  True],
        [ True,  True, False]])
>>> nested_tensor.concat  # Concatenated version (no padding)
tensor([1, 2, 3, 4, 5])
Python Console Session
1
2
3
4
5
6
7
8
9
>>> nested_tensor[0]  # First tensor (no padding)
tensor([1, 2, 3])
>>> nested_tensor[:]  # Padded tensor and mask
(tensor([[1, 2, 3],
        [4, 5, 0]]), tensor([[ True,  True,  True],
        [ True,  True, False]]))
>>> nested_tensor[:, 1:]  # Slice operations return a new NestedTensor
NestedTensor([[2, 3],
        [5, 0]])

Type conversion:

Python Console Session
1
2
3
4
5
6
>>> nested_tensor.to(torch.float).tensor
tensor([[1., 2., 3.],
        [4., 5., 0.]])
>>> nested_tensor.half().tensor
tensor([[1., 2., 3.],
        [4., 5., 0.]], dtype=torch.float16)

Conversion to Python types:

Python Console Session
1
2
>>> nested_tensor.tolist()
[[1, 2, 3], [4, 5]]

Creating from Python lists:

Python Console Session
1
2
3
>>> NestedTensor(*[[1, 2, 3], [4, 5]])
NestedTensor([[1, 2, 3],
        [4, 5, 0]])
Source code in danling/tensor/nested_tensor.py
Python
 175
 176
 177
 178
 179
 180
 181
 182
 183
 184
 185
 186
 187
 188
 189
 190
 191
 192
 193
 194
 195
 196
 197
 198
 199
 200
 201
 202
 203
 204
 205
 206
 207
 208
 209
 210
 211
 212
 213
 214
 215
 216
 217
 218
 219
 220
 221
 222
 223
 224
 225
 226
 227
 228
 229
 230
 231
 232
 233
 234
 235
 236
 237
 238
 239
 240
 241
 242
 243
 244
 245
 246
 247
 248
 249
 250
 251
 252
 253
 254
 255
 256
 257
 258
 259
 260
 261
 262
 263
 264
 265
 266
 267
 268
 269
 270
 271
 272
 273
 274
 275
 276
 277
 278
 279
 280
 281
 282
 283
 284
 285
 286
 287
 288
 289
 290
 291
 292
 293
 294
 295
 296
 297
 298
 299
 300
 301
 302
 303
 304
 305
 306
 307
 308
 309
 310
 311
 312
 313
 314
 315
 316
 317
 318
 319
 320
 321
 322
 323
 324
 325
 326
 327
 328
 329
 330
 331
 332
 333
 334
 335
 336
 337
 338
 339
 340
 341
 342
 343
 344
 345
 346
 347
 348
 349
 350
 351
 352
 353
 354
 355
 356
 357
 358
 359
 360
 361
 362
 363
 364
 365
 366
 367
 368
 369
 370
 371
 372
 373
 374
 375
 376
 377
 378
 379
 380
 381
 382
 383
 384
 385
 386
 387
 388
 389
 390
 391
 392
 393
 394
 395
 396
 397
 398
 399
 400
 401
 402
 403
 404
 405
 406
 407
 408
 409
 410
 411
 412
 413
 414
 415
 416
 417
 418
 419
 420
 421
 422
 423
 424
 425
 426
 427
 428
 429
 430
 431
 432
 433
 434
 435
 436
 437
 438
 439
 440
 441
 442
 443
 444
 445
 446
 447
 448
 449
 450
 451
 452
 453
 454
 455
 456
 457
 458
 459
 460
 461
 462
 463
 464
 465
 466
 467
 468
 469
 470
 471
 472
 473
 474
 475
 476
 477
 478
 479
 480
 481
 482
 483
 484
 485
 486
 487
 488
 489
 490
 491
 492
 493
 494
 495
 496
 497
 498
 499
 500
 501
 502
 503
 504
 505
 506
 507
 508
 509
 510
 511
 512
 513
 514
 515
 516
 517
 518
 519
 520
 521
 522
 523
 524
 525
 526
 527
 528
 529
 530
 531
 532
 533
 534
 535
 536
 537
 538
 539
 540
 541
 542
 543
 544
 545
 546
 547
 548
 549
 550
 551
 552
 553
 554
 555
 556
 557
 558
 559
 560
 561
 562
 563
 564
 565
 566
 567
 568
 569
 570
 571
 572
 573
 574
 575
 576
 577
 578
 579
 580
 581
 582
 583
 584
 585
 586
 587
 588
 589
 590
 591
 592
 593
 594
 595
 596
 597
 598
 599
 600
 601
 602
 603
 604
 605
 606
 607
 608
 609
 610
 611
 612
 613
 614
 615
 616
 617
 618
 619
 620
 621
 622
 623
 624
 625
 626
 627
 628
 629
 630
 631
 632
 633
 634
 635
 636
 637
 638
 639
 640
 641
 642
 643
 644
 645
 646
 647
 648
 649
 650
 651
 652
 653
 654
 655
 656
 657
 658
 659
 660
 661
 662
 663
 664
 665
 666
 667
 668
 669
 670
 671
 672
 673
 674
 675
 676
 677
 678
 679
 680
 681
 682
 683
 684
 685
 686
 687
 688
 689
 690
 691
 692
 693
 694
 695
 696
 697
 698
 699
 700
 701
 702
 703
 704
 705
 706
 707
 708
 709
 710
 711
 712
 713
 714
 715
 716
 717
 718
 719
 720
 721
 722
 723
 724
 725
 726
 727
 728
 729
 730
 731
 732
 733
 734
 735
 736
 737
 738
 739
 740
 741
 742
 743
 744
 745
 746
 747
 748
 749
 750
 751
 752
 753
 754
 755
 756
 757
 758
 759
 760
 761
 762
 763
 764
 765
 766
 767
 768
 769
 770
 771
 772
 773
 774
 775
 776
 777
 778
 779
 780
 781
 782
 783
 784
 785
 786
 787
 788
 789
 790
 791
 792
 793
 794
 795
 796
 797
 798
 799
 800
 801
 802
 803
 804
 805
 806
 807
 808
 809
 810
 811
 812
 813
 814
 815
 816
 817
 818
 819
 820
 821
 822
 823
 824
 825
 826
 827
 828
 829
 830
 831
 832
 833
 834
 835
 836
 837
 838
 839
 840
 841
 842
 843
 844
 845
 846
 847
 848
 849
 850
 851
 852
 853
 854
 855
 856
 857
 858
 859
 860
 861
 862
 863
 864
 865
 866
 867
 868
 869
 870
 871
 872
 873
 874
 875
 876
 877
 878
 879
 880
 881
 882
 883
 884
 885
 886
 887
 888
 889
 890
 891
 892
 893
 894
 895
 896
 897
 898
 899
 900
 901
 902
 903
 904
 905
 906
 907
 908
 909
 910
 911
 912
 913
 914
 915
 916
 917
 918
 919
 920
 921
 922
 923
 924
 925
 926
 927
 928
 929
 930
 931
 932
 933
 934
 935
 936
 937
 938
 939
 940
 941
 942
 943
 944
 945
 946
 947
 948
 949
 950
 951
 952
 953
 954
 955
 956
 957
 958
 959
 960
 961
 962
 963
 964
 965
 966
 967
 968
 969
 970
 971
 972
 973
 974
 975
 976
 977
 978
 979
 980
 981
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
class NestedTensor:
    r"""
    A container for variable-length tensors that enables efficient batch operations.

    `NestedTensor` solves a fundamental problem in deep learning: handling sequences of different lengths
    in batch operations. Instead of excessive padding or complex bucketing, `NestedTensor` provides an
    elegant solution that maintains both efficiency and usability.

    The class provides three main views of the data:
    - `.tensor`: A padded tensor with zeros (or other value) in place of missing elements
    - `.mask`: A boolean mask indicating which elements are real vs padding
    - `.concat`: A flattened tensor containing all elements concatenated (no padding)

    When indexing a `NestedTensor`, the behavior depends on the index type:
    1. Integer index (`nt[0]`): Returns a single tensor without padding
    2. Slice index (`nt[:]`): Returns a tuple of (padded_tensor, mask)
    3. Tuple index (`nt[:, 1:]`): Returns a new `NestedTensor` with the specified sliced shape

    Attributes:
        _storage: The sequence of original tensors (internal use)
        tensor: Padded tensor representing all sequences with padding
        mask: Boolean mask where True indicates real elements, False indicates padding
        concat: Concatenated tensor of all sequences without padding
        batch_first: Whether the first dimension is the batch dimension (B, N, *)
            If `False`, the first dimension is the sequence dimension (N, B, *)
        padding_value: Value used for padding in the padded tensor
        mask_value: Value used in the mask to indicate padding positions (usually False)

    Args:
        *tensors: Variable-length tensors or sequences to store
        batch_first: Whether to use batch-first representation.
        padding_value: Value to use for padding.
        mask_value: Value to use for padding positions in mask.

    Raises:
        ValueError: If `tensors` is not an iterable or is empty

    Examples:
        Basic usage:
        >>> nested_tensor = NestedTensor(torch.tensor([1, 2, 3]), torch.tensor([4, 5]))
        >>> nested_tensor.shape
        torch.Size([2, 3])
        >>> nested_tensor.tensor  # Padded representation
        tensor([[1, 2, 3],
                [4, 5, 0]])
        >>> nested_tensor.mask  # Mask showing real vs padding values
        tensor([[ True,  True,  True],
                [ True,  True, False]])
        >>> nested_tensor.concat  # Concatenated version (no padding)
        tensor([1, 2, 3, 4, 5])

        Indexing:
        >>> nested_tensor[0]  # First tensor (no padding)
        tensor([1, 2, 3])
        >>> nested_tensor[:]  # Padded tensor and mask
        (tensor([[1, 2, 3],
                [4, 5, 0]]), tensor([[ True,  True,  True],
                [ True,  True, False]]))
        >>> nested_tensor[:, 1:]  # Slice operations return a new NestedTensor
        NestedTensor([[2, 3],
                [5, 0]])

        Type conversion:
        >>> nested_tensor.to(torch.float).tensor
        tensor([[1., 2., 3.],
                [4., 5., 0.]])
        >>> nested_tensor.half().tensor
        tensor([[1., 2., 3.],
                [4., 5., 0.]], dtype=torch.float16)

        Conversion to Python types:
        >>> nested_tensor.tolist()
        [[1, 2, 3], [4, 5]]

        Creating from Python lists:
        >>> NestedTensor(*[[1, 2, 3], [4, 5]])
        NestedTensor([[1, 2, 3],
                [4, 5, 0]])
    """

    __storage: Sequence[Tensor]
    batch_first: bool = True
    padding_value: SupportsFloat = 0.0
    mask_value: bool = False

    def __init__(
        self,
        *tensors: Iterable[Tensor],
        batch_first: bool = True,
        padding_value: SupportsFloat = 0.0,
        mask_value: bool = False,
    ) -> None:
        if len(tensors) == 1 and isinstance(tensors, Sequence):
            tensors = tensors[0]  # type: ignore
        self._storage = tensors
        self.batch_first = batch_first
        self.padding_value = padding_value
        self.mask_value = mask_value

    @property
    def _storage(self):
        return self.__storage

    @_storage.setter
    def _storage(self, tensors: Sequence):
        if not isinstance(tensors, Iterable):
            raise ValueError(f"tensors must be an Iterable, bug got {type(tensors)}.")
        tensors = list(tensors)
        if len(tensors) == 0:
            raise ValueError("tensors must be a non-empty Iterable.")
        if not isinstance(tensors[0], Tensor):
            tensors = [torch.tensor(t) for t in tensors]
        # if drop_last=False, the last element is likely not a NestedTensor and has an extra batch dimension
        ndims = {t.ndim for t in tensors[:-1]}
        if len(ndims) == 1:
            (ndim,) = ndims
            if tensors[-1].ndim == ndim + 1 and tensors[-1].size(0) == 1:
                tensors[-1] = tensors[-1].squeeze(0)
        self.__storage = tensors

    def storage(self):
        return self._storage

    @property
    def tensor_mask(self) -> Tuple[Tensor, Tensor]:
        r"""
        Return a tuple of padded tensor and mask tensor.

        Returns:
            (torch.Tensor):

        Examples:
            >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
            >>> nested_tensor.tensor_mask
            (tensor([[1, 2, 3],
                    [4, 5, 0]]), tensor([[ True,  True,  True],
                    [ True,  True, False]]))
        """

        return self._tensor_mask(tuple(self._storage), self.batch_first, self.padding_value, self.mask_value)

    @property
    def tensor(self) -> Tensor:
        r"""
        Return a single tensor by padding all the tensors.

        Returns:
            (torch.Tensor):

        Examples:
            >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
            >>> nested_tensor.tensor
            tensor([[1, 2, 3],
                    [4, 5, 0]])
        """

        return self._tensor(tuple(self._storage), self.batch_first, self.padding_value)

    @property
    def mask(self) -> Tensor:
        r"""
        Padding mask of `tensor`.

        Returns:
            (torch.Tensor):

        Examples:
            >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
            >>> nested_tensor.mask
            tensor([[ True,  True,  True],
                    [ True,  True, False]])
        """

        return self._mask(tuple(self._storage), self.batch_first, self.mask_value)

    @property
    def concat(self) -> Tensor:
        r"""
        Concat `tensor` in padding dim.

        This is particularly useful when calculating loss or passing `Linear` to avoid unnecessary computation.

        Returns:
            (torch.Tensor):

        Examples:
            >>> nested_tensor = NestedTensor([torch.randn(9, 8), torch.randn(11, 8)])
            >>> nested_tensor.concat.shape
            torch.Size([20, 8])
            >>> nested_tensor = NestedTensor([torch.randn(9, 9, 8), torch.randn(11, 11, 8)])
            >>> nested_tensor.concat.shape
            torch.Size([202, 8])
            >>> nested_tensor = NestedTensor([torch.randn(9, 9, 8, 6), torch.randn(11, 11, 8, 6)])
            >>> nested_tensor.concat.shape
            torch.Size([202, 8, 6])
            >>> nested_tensor = NestedTensor([torch.randn(9, 9, 8, 7), torch.randn(11, 11, 8, 6)])
            >>> nested_tensor.concat.shape
            torch.Size([1293, 8])
            >>> nested_tensor = NestedTensor([torch.randn(1, 9, 9, 5), torch.randn(1, 11, 11, 5)])
        """
        shape = list(self.size())  # type: ignore[arg-type]
        shape = shape[1:] if self.batch_first else shape[0] + shape[2:]
        elem = self._storage[0]
        if elem.shape == shape:
            return torch.cat(self._storage, dim=1 if self.batch_first else 0)
        static_dims = set(range(len(shape)))
        for i, s in enumerate(shape):
            if not all(t.size(i) == s for t in self._storage):
                shape[i] = -1
                static_dims.remove(i)
        target_shape = [-1] + [s for s in shape if s != -1]
        storage = [i.reshape(target_shape) for i in self._storage]
        return torch.cat(storage, dim=0 if self.batch_first else 1)

    @classmethod
    def from_tensor_mask(cls, tensor: Tensor, mask: Tensor):
        r"""
        Build a `NestedTensor` object from a padded `Tensor` and corresponding mask `Tensor`.

        Args:
            tensor: Padded Tensor.
            mask: Tensor Mask.

        Returns:
            (torch.Tensor):

        Examples:
            >>> padded_tensor = torch.tensor([[1, 2, 3, 0, 0],
            ...                                [4, 5, 0, 0, 0],
            ...                                [6, 7, 8, 9, 0]])
            >>> mask_tensor = torch.tensor([[1, 1, 1, 0, 0],
            ...                             [1, 1, 0, 0, 0],
            ...                             [1, 1, 1, 1, 0]])
            >>> nested_tensor = NestedTensor.from_tensor_mask(padded_tensor, mask_tensor)
            >>> nested_tensor
            NestedTensor([[1, 2, 3, 0],
                    [4, 5, 0, 0],
                    [6, 7, 8, 9]])
        """

        if mask.ndim == 2:
            return cls(t[slice(0, m.sum())] for t, m in zip(tensor, mask))
        return cls(
            t[[slice(0, (m.sum(dim=dim) > 0).sum().item()) for dim in reversed(range(m.dim()))]]
            for t, m in zip(tensor, mask)
        )

    def nested_like(self, tensor: Tensor, strict: bool = True) -> NestedTensor:
        r"""
        Create a new `NestedTensor` from a `Tensor`.
        The newly created `NestedTensor` will have the same shape as current `NestedTensor`.

        Args:
            tensor: The tensor to be converted to `NestedTensor`.
            strict: Check if the shape of `tensor` is the same as the current `NestedTensor`.

        Returns:
            (NestedTensor):

        Examples:
            >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
            >>> (nested_tensor == nested_tensor.nested_like(nested_tensor)).all()
            tensor(True)
            >>> tensor = nested_tensor.tensor
            >>> (nested_tensor == nested_tensor.nested_like(tensor)).all()
            tensor(True)
            >>> f = nested_tensor.nested_like(torch.randn(2, 2))
            Traceback (most recent call last):
            ValueError: The shape of NestedTensor and input tensor does not match, torch.Size([2, 3]) != torch.Size([2, 2])
            >>> p = nested_tensor.nested_like(torch.randn(2, 2), False)
            >>> p = nested_tensor.nested_like(torch.randn(3, 3), False)
            Traceback (most recent call last):
            ValueError: The batch size of NestedTensor and input tensor does not match, 2 != 3
        """  # noqa: E501

        if isinstance(tensor, NestedTensor):
            return tensor.clone()

        if strict and self.shape != tensor.shape:
            raise ValueError(
                f"The shape of NestedTensor and input tensor does not match, {self.shape} != {tensor.shape}"
            )
        if self.size(0) != tensor.size(0):
            raise ValueError(
                f"The batch size of NestedTensor and input tensor does not match, {self.size(0)} != {tensor.size(0)}"
            )
        return NestedTensor([o[tuple(slice(0, dim) for dim in t.shape)] for t, o in zip(self._storage, tensor)])

    @property
    def device(self) -> torch.device:
        r"""
        Device of the NestedTensor.

        Returns:
            (torch.Tensor):

        Examples:
            >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
            >>> nested_tensor.device
            device(type='cpu')
        """

        return self._device(tuple(self._storage))

    @property
    def shape(self) -> torch.Size | int:
        r"""
        Alias for `size()`.
        """

        return self.size()

    @property
    def ndim(self) -> int:
        r"""
        Alias for `dim()`.
        """

        return self.dim()

    def size(self, dim: int | None = None) -> torch.Size | int:
        r"""
        Returns the size of the self `NestedTensor`.

        Args:
            dim: If not specified, the returned value is a `torch.Size`, a subclass of `tuple`.
                If specified, returns an `int` holding the size of that dimension.
                Defaults to `None`.

        Returns:
            (torch.Size | int):

        Examples:
            >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
            >>> nested_tensor.size()
            torch.Size([2, 3])
            >>> nested_tensor.size(0)
            2
            >>> nested_tensor.storage()[1] = torch.tensor([4, 5, 6, 7])
            >>> nested_tensor.shape
            torch.Size([2, 4])
            >>> nested_tensor.size(1)
            4
        """

        return self._size(tuple(self._storage), dim, self.batch_first)

    def dim(self) -> int:
        r"""
        Number of dimension of the NestedTensor.

        Returns:
            (int):

        Examples:
            >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
            >>> nested_tensor.dim()
            2
            >>> nested_tensor.storage().append(torch.tensor([6, 7, 8, 9]))
            >>> nested_tensor.ndim
            2
        """

        return self._dim(tuple(self._storage))

    def tolist(self) -> list:
        r"""
        Convert a NestedTensor to a list of lists of values.

        Returns:
            (list):

        Examples:
            >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
            >>> nested_tensor.tolist()
            [[1, 2, 3], [4, 5]]
        """

        return [t.tolist() for t in self._storage]

    def all(self, dim: int | None = None, keepdim: bool = False) -> bool | Tensor | NestedTensor:
        r"""
        Tests if all elements in NestedTensor evaluate to True.

        Returns:
            (bool | Tensor):

        Examples:
            >>> nested_tensor = NestedTensor([torch.ones(2, 4, dtype=torch.bool), torch.ones(3, 5, dtype=torch.bool)])
            >>> nested_tensor.all()
            tensor(True)
            >>> nested_tensor.all(dim=0)
            tensor([True, True])
            >>> nested_tensor.all(dim=0, keepdim=True)
            tensor([[True, True]])
            >>> nested_tensor.all(dim=1)
            NestedTensor([[ True,  True,  True,  True, False],
                    [ True,  True,  True,  True,  True]])
            >>> nested_tensor.all(dim=1, keepdim=True)
            NestedTensor([[[ True,  True,  True,  True, False]],
            <BLANKLINE>
                    [[ True,  True,  True,  True,  True]]])
            >>> nested_tensor.batch_first = False
            >>> nested_tensor.all(dim=1)
            tensor([True, True])
            >>> nested_tensor.batch_first = False
            >>> nested_tensor.all(dim=0)
            NestedTensor([[ True,  True,  True,  True, False],
                    [ True,  True,  True,  True,  True]])
            >>> nested_tensor.all(dim=1)
            tensor([True, True])
        """

        if dim is None:
            return torch.tensor(all(i.all() for i in self._storage))
        if (self.batch_first and dim == 0) or (not self.batch_first and dim == 1):
            if keepdim:
                return torch.tensor([i.all() for i in self._storage]).unsqueeze(0 if self.batch_first else 1)
            return torch.tensor([i.all() for i in self._storage])
        if self.batch_first or dim != 0:
            dim -= 1
        return NestedTensor([i.all(dim=dim, keepdim=keepdim) for i in self._storage])

    def where(self, condition: Tensor | NestedTensor, other: Tensor | NestedTensor | SupportsFloat) -> NestedTensor:
        r"""
        Return a NestedTensor of elements selected from either self or other, depending on condition.

        Returns:
            (NestedTensor):

        Examples:
            >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
            >>> nested_tensor.where(nested_tensor > 2, torch.tensor([[6, 5, 4], [3, 2, 1]]))
            NestedTensor([[6, 5, 3],
                    [4, 5, 0]])
            >>> nested_tensor.where(nested_tensor > 2, NestedTensor([[6, 5, 4], [3, 2]]))
            NestedTensor([[6, 5, 3],
                    [4, 5, 0]])
            >>> nested_tensor.where(torch.tensor(True), NestedTensor([[6, 5, 4], [3, 2]]))
            NestedTensor([[1, 2, 3],
                    [4, 5, 0]])
        """

        if isinstance(condition, Tensor) and self.shape == condition.shape:
            condition = self.nested_like(condition)
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(condition, NestedTensor) and isinstance(other, NestedTensor):
            return NestedTensor(
                [x.where(c, y) for x, c, y in zip(self._storage, condition._storage, other._storage)], **self._state
            )
        if isinstance(condition, NestedTensor):
            return NestedTensor([x.where(c, other) for x, c in zip(self._storage, condition._storage)], **self._state)
        if isinstance(other, NestedTensor):
            return NestedTensor([x.where(condition, y) for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor(x.where(condition, other) for x in self._storage)

    @classmethod
    def __torch_function__(cls, func, types, args=(), kwargs=None):
        if kwargs is None:
            kwargs = {}
        if func not in NestedTensorFunc or not all(issubclass(t, (torch.Tensor, NestedTensor)) for t in types):
            args = [a.tensor if hasattr(a, "tensor") else a for a in args]
            return func(*args, **kwargs)
        return NestedTensorFunc[func](*args, **kwargs)

    def __getitem__(self, index: int | slice | list | tuple) -> Tensor | tuple[Tensor, Tensor] | NestedTensor:
        if isinstance(index, int):
            return self._storage[index]
        if isinstance(index, (slice, list)):
            storage = tuple(self._storage[index] if isinstance(index, slice) else [self._storage[i] for i in index])
            pad = pad_tensor(
                storage, size=self._size(storage), batch_first=self.batch_first, padding_value=float(self.padding_value)
            )
            mask = mask_tensor(
                storage, size=self._size(storage), batch_first=self.batch_first, mask_value=self.mask_value
            )
            return pad, mask
        if isinstance(index, tuple):
            return NestedTensor([t[index[0]][index[1:]] for t in self._storage])
        raise ValueError(f"Unsupported index type {type(index)}")

    def __getattr__(self, name: str) -> Any:
        if not self._storage:
            raise ValueError(f"Unable to get {name} from an empty {self.__class__.__name__}")
        ret = [getattr(i, name) for i in self._storage]
        elem = ret[0]
        if isinstance(elem, Tensor):
            return NestedTensor(ret, **self._state)
        if callable(elem):
            return NestedTensorFuncWrapper(ret, state=self._state)
        if elem.__hash__ is not None and len(set(ret)) == 1:
            return elem
        return ret

    @property
    def _state(self) -> Mapping:
        return {k: v for k, v in self.__dict__.items() if not (k.startswith("_") or k.endswith("_"))}

    def __state__(self) -> Mapping:
        return self.__dict__

    def __setstate__(self, state: Mapping) -> None:
        self.__dict__.update(state)

    def __len__(self) -> int:
        return len(self._storage)

    def __repr__(self):
        return self.__class__.__name__ + repr(self.tensor)[len(self.tensor.__class__.__name__) :]  # noqa: E203

    def __bool__(self) -> int:
        return all(bool(x) for x in self._storage)

    def __gt__(  # type: ignore[override]
        self, other: Tensor | NestedTensor | SupportsFloat
    ) -> bool | Tensor | NestedTensor:
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor(i > j for i, j in zip(self._storage, other._storage))
        if isinstance(other, (int, float, Tensor)):
            return NestedTensor([x > other for x in self._storage], **self._state)
        raise TypeError(f"> not supported between instances of '{type(self)}' and '{type(other)}'")

    def __ge__(  # type: ignore[override]
        self, other: Tensor | NestedTensor | SupportsFloat
    ) -> bool | Tensor | NestedTensor:
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor(i >= j for i, j in zip(self._storage, other._storage))
        if isinstance(other, (int, float, Tensor)):
            return NestedTensor([x >= other for x in self._storage], **self._state)
        raise TypeError(f">= not supported between instances of '{type(self)}' and '{type(other)}'")

    def __eq__(  # type: ignore[override]
        self, other: Tensor | NestedTensor | SupportsFloat
    ) -> bool | Tensor | NestedTensor:
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor(i == j for i, j in zip(self._storage, other._storage))
        if isinstance(other, (int, float, Tensor)):
            return NestedTensor([x == other for x in self._storage], **self._state)
        return False

    def __le__(  # type: ignore[override]
        self, other: Tensor | NestedTensor | SupportsFloat
    ) -> bool | Tensor | NestedTensor:
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor(i <= j for i, j in zip(self._storage, other._storage))
        if isinstance(other, (int, float, Tensor)):
            return NestedTensor([x <= other for x in self._storage], **self._state)
        raise TypeError(f"<= not supported between instances of '{type(self)}' and '{type(other)}'")

    def __lt__(  # type: ignore[override]
        self, other: Tensor | NestedTensor | SupportsFloat
    ) -> bool | Tensor | NestedTensor:
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor(i < j for i, j in zip(self._storage, other._storage))
        if isinstance(other, (int, float, Tensor)):
            return NestedTensor([x < other for x in self._storage], **self._state)
        raise TypeError(f"< not supported between instances of '{type(self)}' and '{type(other)}'")

    def __abs__(self):
        return NestedTensor([abs(value) for value in self._storage], **self._state)

    def __add__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([x + y for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([value + other for value in self._storage], **self._state)

    def __radd__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([y + x for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([other + value for value in self._storage], **self._state)

    def __iadd__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if hasattr(other, "to"):
            other = other.to(self.dtype)
        if isinstance(other, NestedTensor):
            for x, y in zip(self._storage, other._storage):
                x += y
        else:
            for value in self._storage:
                value += other
        return self

    def __sub__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([x - y for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([value - other for value in self._storage], **self._state)

    def __rsub__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([y - x for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([other - value for value in self._storage], **self._state)

    def __isub__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if hasattr(other, "to"):
            other = other.to(self.dtype)
        if isinstance(other, NestedTensor):
            for x, y in zip(self._storage, other._storage):
                x -= y
        else:
            for value in self._storage:
                value -= other
        return self

    def __pos__(self):
        return NestedTensor([+x for x in self._storage])

    def __neg__(self):
        return NestedTensor([-x for x in self._storage])

    def __invert__(self):
        return NestedTensor([~x for x in self._storage])

    def __mul__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([x * y for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([value * other for value in self._storage], **self._state)

    def __rmul__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([y * x for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([other * value for value in self._storage], **self._state)

    def __imul__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if hasattr(other, "to"):
            other = other.to(self.dtype)
        if isinstance(other, NestedTensor):
            for x, y in zip(self._storage, other._storage):
                x *= y
        else:
            for value in self._storage:
                value *= other
        return self

    def __pow__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([x**y for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([value**other for value in self._storage], **self._state)

    def __rpow__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([y**x for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([other**value for value in self._storage], **self._state)

    def __ipow__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if hasattr(other, "to"):
            other = other.to(self.dtype)
        if isinstance(other, NestedTensor):
            for x, y in zip(self._storage, other._storage):
                x **= y
        else:
            for value in self._storage:
                value **= other
        return self

    def __matmul__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([x @ y for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([value @ other for value in self._storage], **self._state)

    def __rmatmul__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([y @ x for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([other @ value for value in self._storage], **self._state)

    def __imatmul__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if hasattr(other, "to"):
            other = other.to(self.dtype)
        if isinstance(other, NestedTensor):
            for x, y in zip(self._storage, other._storage):
                x @= y
        else:
            for value in self._storage:
                value @= other
        return self

    def __truediv__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([x / y for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([value / other for value in self._storage], **self._state)

    def __rtruediv__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([y / x for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([other / value for value in self._storage], **self._state)

    def __itruediv__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if hasattr(other, "to"):
            other = other.to(self.dtype)
        if isinstance(other, NestedTensor):
            for x, y in zip(self._storage, other._storage):
                x /= y
        else:
            for value in self._storage:
                value /= other
        return self

    def __floordiv__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([x // y for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([value // other for value in self._storage], **self._state)

    def __rfloordiv__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([y // x for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([other // value for value in self._storage], **self._state)

    def __ifloordiv__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if hasattr(other, "to"):
            other = other.to(self.dtype)
        if isinstance(other, NestedTensor):
            for x, y in zip(self._storage, other._storage):
                x //= y
        else:
            for value in self._storage:
                value //= other
        return self

    def __mod__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([x % y for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([value % other for value in self._storage], **self._state)

    def __rmod__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([y % x for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([other % value for value in self._storage], **self._state)

    def __imod__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if hasattr(other, "to"):
            other = other.to(self.dtype)
        if isinstance(other, NestedTensor):
            for x, y in zip(self._storage, other._storage):
                x %= y
        else:
            for value in self._storage:
                value %= other
        return self

    def __and__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([x & y for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([value & other for value in self._storage], **self._state)

    def __rand__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([y & x for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([other & value for value in self._storage], **self._state)

    def __iand__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if hasattr(other, "to"):
            other = other.to(self.dtype)
        if isinstance(other, NestedTensor):
            for x, y in zip(self._storage, other._storage):
                x &= y
        else:
            for value in self._storage:
                value &= other
        return self

    def __or__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([x | y for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([value | other for value in self._storage], **self._state)

    def __ror__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([y | x for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([other | value for value in self._storage], **self._state)

    def __ior__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if hasattr(other, "to"):
            other = other.to(self.dtype)
        if isinstance(other, NestedTensor):
            for x, y in zip(self._storage, other._storage):
                x |= y
        else:
            for value in self._storage:
                value |= other
        return self

    def __xor__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([x ^ y for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([value ^ other for value in self._storage], **self._state)

    def __rxor__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([y ^ x for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([other ^ value for value in self._storage], **self._state)

    def __ixor__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if hasattr(other, "to"):
            other = other.to(self.dtype)
        if isinstance(other, NestedTensor):
            for x, y in zip(self._storage, other._storage):
                x ^= y
        else:
            for value in self._storage:
                value ^= other
        return self

    def __lshift__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([x << y for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([value << other for value in self._storage], **self._state)

    def __rlshift__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([y << x for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([other << value for value in self._storage], **self._state)

    def __ilshift__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if hasattr(other, "to"):
            other = other.to(self.dtype)
        if isinstance(other, NestedTensor):
            for x, y in zip(self._storage, other._storage):
                x <<= y
        else:
            for value in self._storage:
                value <<= other
        return self

    def __rshift__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([x >> y for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([value >> other for value in self._storage], **self._state)

    def __rrshift__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([y >> x for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([other >> value for value in self._storage], **self._state)

    def __irshift__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if hasattr(other, "to"):
            other = other.to(self.dtype)
        if isinstance(other, NestedTensor):
            for x, y in zip(self._storage, other._storage):
                x >>= y
        else:
            for value in self._storage:
                value >>= other
        return self

    @method_cache(maxsize=1)
    def _tensor_mask(self, storage: tuple, batch_first: bool, padding_value: SupportsFloat, mask_value: bool) -> Tensor:
        if storage[0].dim() == 0:
            return torch.stack(storage, dim=0), torch.full(
                (len(storage),), not mask_value, dtype=torch.bool, device=self.device
            )
        return tensor_mask(
            storage,
            size=self.size(),
            batch_first=batch_first,
            padding_value=float(padding_value),
            mask_value=mask_value,
        )

    @method_cache(maxsize=1)
    def _tensor(self, storage: tuple, batch_first: bool, padding_value: SupportsFloat) -> Tensor:
        if storage[0].dim() == 0:
            return torch.stack(storage, dim=0)
        return pad_tensor(storage, size=self.size(), batch_first=batch_first, padding_value=float(padding_value))

    @method_cache(maxsize=1)
    def _mask(self, storage: tuple, batch_first: bool, mask_value: bool) -> Tensor:
        if storage[0].dim() == 0:
            return torch.full((len(storage),), not mask_value, dtype=torch.bool, device=self.device)
        return mask_tensor(storage, size=self.size(), batch_first=batch_first, mask_value=mask_value)

    @method_cache(maxsize=1)
    def _device(self, storage) -> torch.device:
        return storage[0].device

    @method_cache(maxsize=1)
    def _size(self, storage, dim: int | None = None, batch_first: bool = True) -> torch.Size | int:
        if dim is not None:
            if dim == 0:
                return len(storage)
            return max(t.size(dim - 1) for t in storage)
        if max(t.dim() for t in storage) == 0:
            return torch.Size((len(storage),))
        ndim = max(t.dim() for t in storage)
        size = [max(t.shape[i] if i < len(t.shape) else 0 for t in storage) for i in range(ndim)]
        size.insert(0 if batch_first else 1, len(storage))
        return torch.Size(size)

    @method_cache(maxsize=1)
    def _dim(self, storage) -> torch.Size:
        return max(t.dim() for t in storage) + 1

    def view(self, *shape) -> Tensor:
        r"""
        Returns a torch tensor with a different shape.

        Note:
            since NestedTensor is a collection of tensors, the view operation is ambiguous.

            Therefore, it is converted to a tensor and then reshaped.

        Args:
            shape: The desired size of each dimension.

        Returns:
            (Tensor):

        Examples:
            >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
            >>> nested_tensor.view(3, 2)
            tensor([[1, 2],
                    [3, 4],
                    [5, 0]])
            >>> nested_tensor.view(2, 3)
            tensor([[1, 2, 3],
                    [4, 5, 0]])
        """

        return self.tensor.view(*shape)

    def reshape(self, *shape) -> Tensor:
        r"""
        Returns a torch tensor with a different shape.

        Note:
            since NestedTensor is a collection of tensors, the reshape operation is ambiguous.

            Therefore, it is converted to a tensor and then reshaped.

        Args:
            shape: The desired size of each dimension.

        Returns:
            (Tensor):

        Examples:
            >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
            >>> nested_tensor.reshape(3, 2)
            tensor([[1, 2],
                    [3, 4],
                    [5, 0]])
            >>> nested_tensor.reshape(2, 3)
            tensor([[1, 2, 3],
                    [4, 5, 0]])
        """

        return self.tensor.reshape(*shape)

    def __iter__(self):
        return iter(self._storage)

tensor_mask property

Python
tensor_mask: Tuple[Tensor, Tensor]

Return a tuple of padded tensor and mask tensor.

Returns:

Type Description
Tensor

Examples:

Python Console Session
1
2
3
4
5
>>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
>>> nested_tensor.tensor_mask
(tensor([[1, 2, 3],
        [4, 5, 0]]), tensor([[ True,  True,  True],
        [ True,  True, False]]))

tensor property

Python
tensor: Tensor

Return a single tensor by padding all the tensors.

Returns:

Type Description
Tensor

Examples:

Python Console Session
1
2
3
4
>>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
>>> nested_tensor.tensor
tensor([[1, 2, 3],
        [4, 5, 0]])

mask property

Python
mask: Tensor

Padding mask of tensor.

Returns:

Type Description
Tensor

Examples:

Python Console Session
1
2
3
4
>>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
>>> nested_tensor.mask
tensor([[ True,  True,  True],
        [ True,  True, False]])

concat property

Python
concat: Tensor

Concat tensor in padding dim.

This is particularly useful when calculating loss or passing Linear to avoid unnecessary computation.

Returns:

Type Description
Tensor

Examples:

Python Console Session
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
>>> nested_tensor = NestedTensor([torch.randn(9, 8), torch.randn(11, 8)])
>>> nested_tensor.concat.shape
torch.Size([20, 8])
>>> nested_tensor = NestedTensor([torch.randn(9, 9, 8), torch.randn(11, 11, 8)])
>>> nested_tensor.concat.shape
torch.Size([202, 8])
>>> nested_tensor = NestedTensor([torch.randn(9, 9, 8, 6), torch.randn(11, 11, 8, 6)])
>>> nested_tensor.concat.shape
torch.Size([202, 8, 6])
>>> nested_tensor = NestedTensor([torch.randn(9, 9, 8, 7), torch.randn(11, 11, 8, 6)])
>>> nested_tensor.concat.shape
torch.Size([1293, 8])
>>> nested_tensor = NestedTensor([torch.randn(1, 9, 9, 5), torch.randn(1, 11, 11, 5)])

device property

Python
device: device

Device of the NestedTensor.

Returns:

Type Description
Tensor

Examples:

Python Console Session
1
2
3
>>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
>>> nested_tensor.device
device(type='cpu')

shape property

Python
shape: Size | int

Alias for size().

ndim property

Python
ndim: int

Alias for dim().

from_tensor_mask classmethod

Python
from_tensor_mask(tensor: Tensor, mask: Tensor)

Build a NestedTensor object from a padded Tensor and corresponding mask Tensor.

Parameters:

Name Type Description Default
tensor
Tensor

Padded Tensor.

 required
mask
Tensor

Tensor Mask.

 required

Returns:

Type Description
Tensor

Examples:

Python Console Session
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
>>> padded_tensor = torch.tensor([[1, 2, 3, 0, 0],
...                                [4, 5, 0, 0, 0],
...                                [6, 7, 8, 9, 0]])
>>> mask_tensor = torch.tensor([[1, 1, 1, 0, 0],
...                             [1, 1, 0, 0, 0],
...                             [1, 1, 1, 1, 0]])
>>> nested_tensor = NestedTensor.from_tensor_mask(padded_tensor, mask_tensor)
>>> nested_tensor
NestedTensor([[1, 2, 3, 0],
        [4, 5, 0, 0],
        [6, 7, 8, 9]])
Source code in danling/tensor/nested_tensor.py
Python
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
@classmethod
def from_tensor_mask(cls, tensor: Tensor, mask: Tensor):
    r"""
    Build a `NestedTensor` object from a padded `Tensor` and corresponding mask `Tensor`.

    Args:
        tensor: Padded Tensor.
        mask: Tensor Mask.

    Returns:
        (torch.Tensor):

    Examples:
        >>> padded_tensor = torch.tensor([[1, 2, 3, 0, 0],
        ...                                [4, 5, 0, 0, 0],
        ...                                [6, 7, 8, 9, 0]])
        >>> mask_tensor = torch.tensor([[1, 1, 1, 0, 0],
        ...                             [1, 1, 0, 0, 0],
        ...                             [1, 1, 1, 1, 0]])
        >>> nested_tensor = NestedTensor.from_tensor_mask(padded_tensor, mask_tensor)
        >>> nested_tensor
        NestedTensor([[1, 2, 3, 0],
                [4, 5, 0, 0],
                [6, 7, 8, 9]])
    """

    if mask.ndim == 2:
        return cls(t[slice(0, m.sum())] for t, m in zip(tensor, mask))
    return cls(
        t[[slice(0, (m.sum(dim=dim) > 0).sum().item()) for dim in reversed(range(m.dim()))]]
        for t, m in zip(tensor, mask)
    )

nested_like

Python
nested_like(
    tensor: Tensor, strict: bool = True
) -> NestedTensor

Create a new NestedTensor from a Tensor. The newly created NestedTensor will have the same shape as current NestedTensor.

Parameters:

Name Type Description Default
tensor
Tensor

The tensor to be converted to NestedTensor.

 required
strict
bool

Check if the shape of tensor is the same as the current NestedTensor.

 True

Returns:

Type Description
NestedTensor

Examples:

Python Console Session
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
>>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
>>> (nested_tensor == nested_tensor.nested_like(nested_tensor)).all()
tensor(True)
>>> tensor = nested_tensor.tensor
>>> (nested_tensor == nested_tensor.nested_like(tensor)).all()
tensor(True)
>>> f = nested_tensor.nested_like(torch.randn(2, 2))
Traceback (most recent call last):
ValueError: The shape of NestedTensor and input tensor does not match, torch.Size([2, 3]) != torch.Size([2, 2])
>>> p = nested_tensor.nested_like(torch.randn(2, 2), False)
>>> p = nested_tensor.nested_like(torch.randn(3, 3), False)
Traceback (most recent call last):
ValueError: The batch size of NestedTensor and input tensor does not match, 2 != 3
Source code in danling/tensor/nested_tensor.py
Python
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
def nested_like(self, tensor: Tensor, strict: bool = True) -> NestedTensor:
    r"""
    Create a new `NestedTensor` from a `Tensor`.
    The newly created `NestedTensor` will have the same shape as current `NestedTensor`.

    Args:
        tensor: The tensor to be converted to `NestedTensor`.
        strict: Check if the shape of `tensor` is the same as the current `NestedTensor`.

    Returns:
        (NestedTensor):

    Examples:
        >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
        >>> (nested_tensor == nested_tensor.nested_like(nested_tensor)).all()
        tensor(True)
        >>> tensor = nested_tensor.tensor
        >>> (nested_tensor == nested_tensor.nested_like(tensor)).all()
        tensor(True)
        >>> f = nested_tensor.nested_like(torch.randn(2, 2))
        Traceback (most recent call last):
        ValueError: The shape of NestedTensor and input tensor does not match, torch.Size([2, 3]) != torch.Size([2, 2])
        >>> p = nested_tensor.nested_like(torch.randn(2, 2), False)
        >>> p = nested_tensor.nested_like(torch.randn(3, 3), False)
        Traceback (most recent call last):
        ValueError: The batch size of NestedTensor and input tensor does not match, 2 != 3
    """  # noqa: E501

    if isinstance(tensor, NestedTensor):
        return tensor.clone()

    if strict and self.shape != tensor.shape:
        raise ValueError(
            f"The shape of NestedTensor and input tensor does not match, {self.shape} != {tensor.shape}"
        )
    if self.size(0) != tensor.size(0):
        raise ValueError(
            f"The batch size of NestedTensor and input tensor does not match, {self.size(0)} != {tensor.size(0)}"
        )
    return NestedTensor([o[tuple(slice(0, dim) for dim in t.shape)] for t, o in zip(self._storage, tensor)])

size

Python
size(dim: int | None = None) -> Size | int

Returns the size of the self NestedTensor.

Parameters:

Name Type Description Default
dim
int | None

If not specified, the returned value is a torch.Size, a subclass of tuple. If specified, returns an int holding the size of that dimension. Defaults to None.

 None

Returns:

Type Description
Size | int

Examples:

Python Console Session
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
>>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
>>> nested_tensor.size()
torch.Size([2, 3])
>>> nested_tensor.size(0)
2
>>> nested_tensor.storage()[1] = torch.tensor([4, 5, 6, 7])
>>> nested_tensor.shape
torch.Size([2, 4])
>>> nested_tensor.size(1)
4
Source code in danling/tensor/nested_tensor.py
Python
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
def size(self, dim: int | None = None) -> torch.Size | int:
    r"""
    Returns the size of the self `NestedTensor`.

    Args:
        dim: If not specified, the returned value is a `torch.Size`, a subclass of `tuple`.
            If specified, returns an `int` holding the size of that dimension.
            Defaults to `None`.

    Returns:
        (torch.Size | int):

    Examples:
        >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
        >>> nested_tensor.size()
        torch.Size([2, 3])
        >>> nested_tensor.size(0)
        2
        >>> nested_tensor.storage()[1] = torch.tensor([4, 5, 6, 7])
        >>> nested_tensor.shape
        torch.Size([2, 4])
        >>> nested_tensor.size(1)
        4
    """

    return self._size(tuple(self._storage), dim, self.batch_first)

dim

Python
dim() -> int

Number of dimension of the NestedTensor.

Returns:

Type Description
int

Examples:

Python Console Session
1
2
3
4
5
6
>>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
>>> nested_tensor.dim()
2
>>> nested_tensor.storage().append(torch.tensor([6, 7, 8, 9]))
>>> nested_tensor.ndim
2
Source code in danling/tensor/nested_tensor.py
Python
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
def dim(self) -> int:
    r"""
    Number of dimension of the NestedTensor.

    Returns:
        (int):

    Examples:
        >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
        >>> nested_tensor.dim()
        2
        >>> nested_tensor.storage().append(torch.tensor([6, 7, 8, 9]))
        >>> nested_tensor.ndim
        2
    """

    return self._dim(tuple(self._storage))

tolist

Python
tolist() -> list

Convert a NestedTensor to a list of lists of values.

Returns:

Type Description
list

Examples:

Python Console Session
1
2
3
>>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
>>> nested_tensor.tolist()
[[1, 2, 3], [4, 5]]
Source code in danling/tensor/nested_tensor.py
Python
540
541
542
543
544
545
546
547
548
549
550
551
552
553
def tolist(self) -> list:
    r"""
    Convert a NestedTensor to a list of lists of values.

    Returns:
        (list):

    Examples:
        >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
        >>> nested_tensor.tolist()
        [[1, 2, 3], [4, 5]]
    """

    return [t.tolist() for t in self._storage]

all

Python
all(
    dim: int | None = None, keepdim: bool = False
) -> bool | Tensor | NestedTensor

Tests if all elements in NestedTensor evaluate to True.

Returns:

Type Description
bool | Tensor

Examples:

Python Console Session
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
>>> nested_tensor = NestedTensor([torch.ones(2, 4, dtype=torch.bool), torch.ones(3, 5, dtype=torch.bool)])
>>> nested_tensor.all()
tensor(True)
>>> nested_tensor.all(dim=0)
tensor([True, True])
>>> nested_tensor.all(dim=0, keepdim=True)
tensor([[True, True]])
>>> nested_tensor.all(dim=1)
NestedTensor([[ True,  True,  True,  True, False],
        [ True,  True,  True,  True,  True]])
>>> nested_tensor.all(dim=1, keepdim=True)
NestedTensor([[[ True,  True,  True,  True, False]],

        [[ True,  True,  True,  True,  True]]])
>>> nested_tensor.batch_first = False
>>> nested_tensor.all(dim=1)
tensor([True, True])
>>> nested_tensor.batch_first = False
>>> nested_tensor.all(dim=0)
NestedTensor([[ True,  True,  True,  True, False],
        [ True,  True,  True,  True,  True]])
>>> nested_tensor.all(dim=1)
tensor([True, True])
Source code in danling/tensor/nested_tensor.py
Python
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
def all(self, dim: int | None = None, keepdim: bool = False) -> bool | Tensor | NestedTensor:
    r"""
    Tests if all elements in NestedTensor evaluate to True.

    Returns:
        (bool | Tensor):

    Examples:
        >>> nested_tensor = NestedTensor([torch.ones(2, 4, dtype=torch.bool), torch.ones(3, 5, dtype=torch.bool)])
        >>> nested_tensor.all()
        tensor(True)
        >>> nested_tensor.all(dim=0)
        tensor([True, True])
        >>> nested_tensor.all(dim=0, keepdim=True)
        tensor([[True, True]])
        >>> nested_tensor.all(dim=1)
        NestedTensor([[ True,  True,  True,  True, False],
                [ True,  True,  True,  True,  True]])
        >>> nested_tensor.all(dim=1, keepdim=True)
        NestedTensor([[[ True,  True,  True,  True, False]],
        <BLANKLINE>
                [[ True,  True,  True,  True,  True]]])
        >>> nested_tensor.batch_first = False
        >>> nested_tensor.all(dim=1)
        tensor([True, True])
        >>> nested_tensor.batch_first = False
        >>> nested_tensor.all(dim=0)
        NestedTensor([[ True,  True,  True,  True, False],
                [ True,  True,  True,  True,  True]])
        >>> nested_tensor.all(dim=1)
        tensor([True, True])
    """

    if dim is None:
        return torch.tensor(all(i.all() for i in self._storage))
    if (self.batch_first and dim == 0) or (not self.batch_first and dim == 1):
        if keepdim:
            return torch.tensor([i.all() for i in self._storage]).unsqueeze(0 if self.batch_first else 1)
        return torch.tensor([i.all() for i in self._storage])
    if self.batch_first or dim != 0:
        dim -= 1
    return NestedTensor([i.all(dim=dim, keepdim=keepdim) for i in self._storage])

where

Python
where(
    condition: Tensor | NestedTensor,
    other: Tensor | NestedTensor | SupportsFloat,
) -> NestedTensor

Return a NestedTensor of elements selected from either self or other, depending on condition.

Returns:

Type Description
NestedTensor

Examples:

Python Console Session
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
>>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
>>> nested_tensor.where(nested_tensor > 2, torch.tensor([[6, 5, 4], [3, 2, 1]]))
NestedTensor([[6, 5, 3],
        [4, 5, 0]])
>>> nested_tensor.where(nested_tensor > 2, NestedTensor([[6, 5, 4], [3, 2]]))
NestedTensor([[6, 5, 3],
        [4, 5, 0]])
>>> nested_tensor.where(torch.tensor(True), NestedTensor([[6, 5, 4], [3, 2]]))
NestedTensor([[1, 2, 3],
        [4, 5, 0]])
Source code in danling/tensor/nested_tensor.py
Python
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
def where(self, condition: Tensor | NestedTensor, other: Tensor | NestedTensor | SupportsFloat) -> NestedTensor:
    r"""
    Return a NestedTensor of elements selected from either self or other, depending on condition.

    Returns:
        (NestedTensor):

    Examples:
        >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
        >>> nested_tensor.where(nested_tensor > 2, torch.tensor([[6, 5, 4], [3, 2, 1]]))
        NestedTensor([[6, 5, 3],
                [4, 5, 0]])
        >>> nested_tensor.where(nested_tensor > 2, NestedTensor([[6, 5, 4], [3, 2]]))
        NestedTensor([[6, 5, 3],
                [4, 5, 0]])
        >>> nested_tensor.where(torch.tensor(True), NestedTensor([[6, 5, 4], [3, 2]]))
        NestedTensor([[1, 2, 3],
                [4, 5, 0]])
    """

    if isinstance(condition, Tensor) and self.shape == condition.shape:
        condition = self.nested_like(condition)
    if isinstance(other, Tensor) and self.shape == other.shape:
        other = self.nested_like(other)
    if isinstance(condition, NestedTensor) and isinstance(other, NestedTensor):
        return NestedTensor(
            [x.where(c, y) for x, c, y in zip(self._storage, condition._storage, other._storage)], **self._state
        )
    if isinstance(condition, NestedTensor):
        return NestedTensor([x.where(c, other) for x, c in zip(self._storage, condition._storage)], **self._state)
    if isinstance(other, NestedTensor):
        return NestedTensor([x.where(condition, y) for x, y in zip(self._storage, other._storage)], **self._state)
    return NestedTensor(x.where(condition, other) for x in self._storage)

view

Python
view(*shape) -> Tensor

Returns a torch tensor with a different shape.

Note

since NestedTensor is a collection of tensors, the view operation is ambiguous.

Therefore, it is converted to a tensor and then reshaped.

Parameters:

Name Type Description Default
shape

The desired size of each dimension.

 ()

Returns:

Type Description
Tensor

Examples:

Python Console Session
1
2
3
4
5
6
7
8
>>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
>>> nested_tensor.view(3, 2)
tensor([[1, 2],
        [3, 4],
        [5, 0]])
>>> nested_tensor.view(2, 3)
tensor([[1, 2, 3],
        [4, 5, 0]])
Source code in danling/tensor/nested_tensor.py
Python
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
def view(self, *shape) -> Tensor:
    r"""
    Returns a torch tensor with a different shape.

    Note:
        since NestedTensor is a collection of tensors, the view operation is ambiguous.

        Therefore, it is converted to a tensor and then reshaped.

    Args:
        shape: The desired size of each dimension.

    Returns:
        (Tensor):

    Examples:
        >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
        >>> nested_tensor.view(3, 2)
        tensor([[1, 2],
                [3, 4],
                [5, 0]])
        >>> nested_tensor.view(2, 3)
        tensor([[1, 2, 3],
                [4, 5, 0]])
    """

    return self.tensor.view(*shape)

reshape

Python
reshape(*shape) -> Tensor

Returns a torch tensor with a different shape.

Note

since NestedTensor is a collection of tensors, the reshape operation is ambiguous.

Therefore, it is converted to a tensor and then reshaped.

Parameters:

Name Type Description Default
shape

The desired size of each dimension.

 ()

Returns:

Type Description
Tensor

Examples:

Python Console Session
1
2
3
4
5
6
7
8
>>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
>>> nested_tensor.reshape(3, 2)
tensor([[1, 2],
        [3, 4],
        [5, 0]])
>>> nested_tensor.reshape(2, 3)
tensor([[1, 2, 3],
        [4, 5, 0]])
Source code in danling/tensor/nested_tensor.py
Python
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
def reshape(self, *shape) -> Tensor:
    r"""
    Returns a torch tensor with a different shape.

    Note:
        since NestedTensor is a collection of tensors, the reshape operation is ambiguous.

        Therefore, it is converted to a tensor and then reshaped.

    Args:
        shape: The desired size of each dimension.

    Returns:
        (Tensor):

    Examples:
        >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
        >>> nested_tensor.reshape(3, 2)
        tensor([[1, 2],
                [3, 4],
                [5, 0]])
        >>> nested_tensor.reshape(2, 3)
        tensor([[1, 2, 3],
                [4, 5, 0]])
    """

    return self.tensor.reshape(*shape)

PNTensor

Bases: Tensor

A tensor wrapper that enables automatic collation into NestedTensor with PyTorch DataLoader.

PNTensor (Potential Nested Tensor) seamlessly bridges the gap between individual tensors and batched NestedTensor objects in PyTorch workflows. It’s designed specifically to work with PyTorch’s DataLoader collation mechanism, allowing datasets to return variable-length tensors that will automatically be combined into a NestedTensor when batched.

The class provides three properties that mirror those of NestedTensor: - .tensor: The tensor itself (self) - .mask: A tensor of ones with the same shape as self - .concat: The tensor itself (self)

Examples:

Basic usage with PyTorch DataLoader:

Python Console Session
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
>>> from torch.utils.data import Dataset, DataLoader
>>> from danling.tensor import PNTensor
>>> class VariableLengthDataset(Dataset):
...     def __init__(self, data):
...         self.data = data
...     def __len__(self):
...         return len(self.data)
...     def __getitem__(self, idx):
...         return PNTensor(self.data[idx])
>>> # Create a dataset with variable-length sequences
>>> dataset = VariableLengthDataset([[1, 2, 3], [4, 5], [6, 7, 8, 9]])
>>> dataloader = DataLoader(dataset, batch_size=3)
>>> # The DataLoader automatically produces NestedTensor batches
>>> batch = next(iter(dataloader))
>>> batch
NestedTensor([[1., 2., 3., 0.],
        [4., 5., 0., 0.],
        [6., 7., 8., 9.]])

Using PNTensor directly:

Python Console Session
1
2
3
4
5
6
7
>>> tensor = PNTensor([1, 2, 3])
>>> tensor
PNTensor([1., 2., 3.])
>>> tensor.tensor
PNTensor([1., 2., 3.])
>>> tensor.mask
PNTensor([True, True, True])
Source code in danling/tensor/nested_tensor.py
Python
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
class PNTensor(Tensor):
    r"""
    A tensor wrapper that enables automatic collation into NestedTensor with PyTorch DataLoader.

    `PNTensor` (Potential Nested Tensor) seamlessly bridges the gap between individual tensors
    and batched `NestedTensor` objects in PyTorch workflows. It's designed specifically to work
    with PyTorch's DataLoader collation mechanism, allowing datasets to return variable-length
    tensors that will automatically be combined into a `NestedTensor` when batched.

    The class provides three properties that mirror those of NestedTensor:
    - `.tensor`: The tensor itself (self)
    - `.mask`: A tensor of ones with the same shape as self
    - `.concat`: The tensor itself (self)

    Attributes:
        Inherits all attributes from torch.Tensor

    Methods:
        Inherits all methods from torch.Tensor

    Examples:
        Basic usage with PyTorch DataLoader:

        >>> from torch.utils.data import Dataset, DataLoader
        >>> from danling.tensor import PNTensor
        >>> class VariableLengthDataset(Dataset):
        ...     def __init__(self, data):
        ...         self.data = data
        ...     def __len__(self):
        ...         return len(self.data)
        ...     def __getitem__(self, idx):
        ...         return PNTensor(self.data[idx])
        >>> # Create a dataset with variable-length sequences
        >>> dataset = VariableLengthDataset([[1, 2, 3], [4, 5], [6, 7, 8, 9]])
        >>> dataloader = DataLoader(dataset, batch_size=3)
        >>> # The DataLoader automatically produces NestedTensor batches
        >>> batch = next(iter(dataloader))
        >>> batch
        NestedTensor([[1., 2., 3., 0.],
                [4., 5., 0., 0.],
                [6., 7., 8., 9.]])

        Using PNTensor directly:

        >>> tensor = PNTensor([1, 2, 3])
        >>> tensor
        PNTensor([1., 2., 3.])
        >>> tensor.tensor
        PNTensor([1., 2., 3.])
        >>> tensor.mask
        PNTensor([True, True, True])
    """

    @property
    def tensor(self) -> Tensor:
        r"""
        Identical to `self`.

        Returns:
            (torch.Tensor):

        Examples:
            >>> tensor = torch.tensor([1, 2, 3])
            >>> pn_tensor = PNTensor([1, 2, 3])
            >>> bool((tensor == pn_tensor).all())
            True
            >>> bool((tensor == pn_tensor.tensor).all())
            True
        """

        return self

    @property
    def mask(self) -> Tensor:
        r"""
        Identical to `torch.ones_like(self)`.

        Returns:
            (torch.Tensor):

        Examples:
            >>> tensor = torch.tensor([1, 2, 3])
            >>> pn_tensor = PNTensor([1, 2, 3])
            >>> bool((pn_tensor.mask == torch.ones_like(pn_tensor)).all().item())
            True
        """

        return torch.ones_like(self, dtype=torch.bool)

    @property
    def contact(self) -> Tensor:
        r"""
        Identical to `self`.

        Returns:
            (torch.Tensor):

        Examples:
            >>> tensor = torch.tensor([1, 2, 3])
            >>> pn_tensor = PNTensor([1, 2, 3])
            >>> bool((tensor == pn_tensor).all())
            True
            >>> bool((tensor == pn_tensor.contact).all())
            True
        """

        return self

    def new_empty(self, *args, **kwargs):
        return PNTensor(super().new_empty(*args, **kwargs))

tensor property

Python
tensor: Tensor

Identical to self.

Returns:

Type Description
Tensor

Examples:

Python Console Session
1
2
3
4
5
6
>>> tensor = torch.tensor([1, 2, 3])
>>> pn_tensor = PNTensor([1, 2, 3])
>>> bool((tensor == pn_tensor).all())
True
>>> bool((tensor == pn_tensor.tensor).all())
True

mask property

Python
mask: Tensor

Identical to torch.ones_like(self).

Returns:

Type Description
Tensor

Examples:

Python Console Session
1
2
3
4
>>> tensor = torch.tensor([1, 2, 3])
>>> pn_tensor = PNTensor([1, 2, 3])
>>> bool((pn_tensor.mask == torch.ones_like(pn_tensor)).all().item())
True

contact property

Python
contact: Tensor

Identical to self.

Returns:

Type Description
Tensor

Examples:

Python Console Session
1
2
3
4
5
6
>>> tensor = torch.tensor([1, 2, 3])
>>> pn_tensor = PNTensor([1, 2, 3])
>>> bool((tensor == pn_tensor).all())
True
>>> bool((tensor == pn_tensor.contact).all())
True

ensure_dir

Bases: property

Ensure a directory property exists.

Examples:

Python Console Session
1
2
3
>>> @ensure_dir
... def dir(self) -> str:
...     return os.path.join("path", "to", "dir")
Source code in danling/utils/descriptors.py
Python
80
81
82
83
84
85
86
87
88
89
90
91
92
93
class ensure_dir(property):
    r"""
    Ensure a directory property exists.

    Examples:
        >>> @ensure_dir
        ... def dir(self) -> str:
        ...     return os.path.join("path", "to", "dir")
    """

    def __get__(self, instance, owner=None):
        val = super().__get__(instance, owner)
        makedirs(val, exist_ok=True)
        return val

Metrics

Bases: Metric

A comprehensive metric tracking system that maintains the complete history of predictions and labels.

Metrics is designed for computing evaluations that require access to the entire dataset history, such as AUROC, Pearson correlation, or other metrics that cannot be meaningfully averaged batch-by-batch.

Attributes:

Name Type Description
metrics FlatDict[str, Callable]

A dictionary of metric functions to be computed
preprocess Optional[Callable]

Optional preprocessing function to apply to inputs and targets
ignore_index int

Value to ignore in classification tasks (e.g., -100 for padding)
ignore_nan bool

Whether to ignore NaN values in regression tasks
val NestedDict[str, float | flist]

Metrics computed on the current batch only
avg NestedDict[str, float | flist]

Metrics computed on all accumulated data
input Tensor

The input tensor from the latest batch
target Tensor

The target tensor from the latest batch
inputs Tensor

Concatenation of all input tensors seen so far
targets Tensor

Concatenation of all target tensors seen so far

Parameters:

Name Type Description Default

*args

A single mapping of metrics or callable metric functions

 ()

device

device | None

Device to store tensors on

 None

ignore_index

int | None

Value to ignore in classification tasks

 None

ignore_nan

bool | None

Whether to ignore NaN values in regression tasks

 None

preprocess

Callable | None

Function to preprocess inputs before computing metrics

 None

**metrics

Callable

Named metric functions to compute

 {}

Examples:

Python Console Session
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
>>> from danling.metric.functional import auroc, auprc, base_preprocess
>>> metrics = Metrics(auroc=auroc, auprc=auprc)
>>> metrics
Metrics('auroc', 'auprc')
>>> metrics.update([0.2, 0.3, 0.5, 0.7], [0, 1, 0, 1])
>>> metrics.input  # predicted values of current batch
tensor([0.2000, 0.3000, 0.5000, 0.7000])
>>> metrics.target  # ground truth of current batch
tensor([0, 1, 0, 1])
>>> metrics.inputs  # predicted values of all data
tensor([0.2000, 0.3000, 0.5000, 0.7000])
>>> metrics.targets  # ground truth of all data
tensor([0, 1, 0, 1])
>>> metrics.val  # Metrics of current batch on current device
NestedDict(
  ('auroc'): 0.75
  ('auprc'): 0.8333333730697632
)
>>> metrics.avg  # Metrics of all data on all devices
NestedDict(
  ('auroc'): 0.75
  ('auprc'): 0.8333333730697632
)
>>> metrics.update([0.1, 0.4, 0.6, 0.8], [0, 0, 1, 0])
>>> metrics.input  # predicted values of current batch
tensor([0.1000, 0.4000, 0.6000, 0.8000])
>>> metrics.target  # ground truth of current batch
tensor([0, 0, 1, 0])
>>> metrics.inputs  # predicted values of all data
tensor([0.2000, 0.3000, 0.5000, 0.7000, 0.1000, 0.4000, 0.6000, 0.8000])
>>> metrics.targets  # ground truth of all data
tensor([0, 1, 0, 1, 0, 0, 1, 0])
>>> metrics.val  # Metrics of current batch on current device
NestedDict(
  ('auroc'): 0.6666666666666666
  ('auprc'): 0.5
)
>>> metrics.avg  # Metrics of all data on all devices
NestedDict(
  ('auroc'): 0.6666666666666666
  ('auprc'): 0.5555555820465088
)
>>> f"{metrics:.4f}"
'auroc: 0.6667 (0.6667)\tauprc: 0.5000 (0.5556)'
>>> metrics = Metrics(auroc=auroc, auprc=auprc, preprocess=base_preprocess)
>>> metrics.update([[0.1, 0.4, 0.6, 0.8], [0.1, 0.4, 0.6]], [[0, -100, 1, 0], [0, -100, 1]])
>>> metrics.input, metrics.target
(tensor([0.1000, 0.6000, 0.8000, 0.1000, 0.6000]), tensor([0, 1, 0, 0, 1]))
Notes
  • Metrics stores the complete prediction and target history, which is memory-intensive but necessary for metrics like AUROC that operate on the entire dataset.
  • For metrics that can be meaningfully averaged batch-by-batch (like accuracy), consider using MetricMeter for better memory efficiency.
  • The ignore_index parameter is useful for handling masked or padded values in classification tasks.
  • The ignore_nan parameter helps handle missing values in regression tasks.
  • All metrics are synchronized across devices in distributed training environments.
See Also
  • MetricMeters: Memory-efficient metric tracker that averages multiple metrics batch-by-batch.
Source code in danling/metric/metrics.py
Python
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
class Metrics(Metric):
    r"""
    A comprehensive metric tracking system that maintains the complete history of predictions and labels.

    Metrics is designed for computing evaluations that require access to the entire dataset history,
    such as AUROC, Pearson correlation, or other metrics that cannot be meaningfully averaged batch-by-batch.

    Attributes:
        metrics: A dictionary of metric functions to be computed
        preprocess: Optional preprocessing function to apply to inputs and targets
        ignore_index: Value to ignore in classification tasks (e.g., -100 for padding)
        ignore_nan: Whether to ignore NaN values in regression tasks
        val: Metrics computed on the current batch only
        avg: Metrics computed on all accumulated data
        input: The input tensor from the latest batch
        target: The target tensor from the latest batch
        inputs: Concatenation of all input tensors seen so far
        targets: Concatenation of all target tensors seen so far

    Args:
        *args: A single mapping of metrics or callable metric functions
        device: Device to store tensors on
        ignore_index: Value to ignore in classification tasks
        ignore_nan: Whether to ignore NaN values in regression tasks
        preprocess: Function to preprocess inputs before computing metrics
        **metrics: Named metric functions to compute

    Examples:
        >>> from danling.metric.functional import auroc, auprc, base_preprocess
        >>> metrics = Metrics(auroc=auroc, auprc=auprc)
        >>> metrics
        Metrics('auroc', 'auprc')
        >>> metrics.update([0.2, 0.3, 0.5, 0.7], [0, 1, 0, 1])
        >>> metrics.input  # predicted values of current batch
        tensor([0.2000, 0.3000, 0.5000, 0.7000])
        >>> metrics.target  # ground truth of current batch
        tensor([0, 1, 0, 1])
        >>> metrics.inputs  # predicted values of all data
        tensor([0.2000, 0.3000, 0.5000, 0.7000])
        >>> metrics.targets  # ground truth of all data
        tensor([0, 1, 0, 1])
        >>> metrics.val  # Metrics of current batch on current device
        NestedDict(
          ('auroc'): 0.75
          ('auprc'): 0.8333333730697632
        )
        >>> metrics.avg  # Metrics of all data on all devices
        NestedDict(
          ('auroc'): 0.75
          ('auprc'): 0.8333333730697632
        )
        >>> metrics.update([0.1, 0.4, 0.6, 0.8], [0, 0, 1, 0])
        >>> metrics.input  # predicted values of current batch
        tensor([0.1000, 0.4000, 0.6000, 0.8000])
        >>> metrics.target  # ground truth of current batch
        tensor([0, 0, 1, 0])
        >>> metrics.inputs  # predicted values of all data
        tensor([0.2000, 0.3000, 0.5000, 0.7000, 0.1000, 0.4000, 0.6000, 0.8000])
        >>> metrics.targets  # ground truth of all data
        tensor([0, 1, 0, 1, 0, 0, 1, 0])
        >>> metrics.val  # Metrics of current batch on current device
        NestedDict(
          ('auroc'): 0.6666666666666666
          ('auprc'): 0.5
        )
        >>> metrics.avg  # Metrics of all data on all devices
        NestedDict(
          ('auroc'): 0.6666666666666666
          ('auprc'): 0.5555555820465088
        )
        >>> f"{metrics:.4f}"
        'auroc: 0.6667 (0.6667)\tauprc: 0.5000 (0.5556)'
        >>> metrics = Metrics(auroc=auroc, auprc=auprc, preprocess=base_preprocess)
        >>> metrics.update([[0.1, 0.4, 0.6, 0.8], [0.1, 0.4, 0.6]], [[0, -100, 1, 0], [0, -100, 1]])
        >>> metrics.input, metrics.target
        (tensor([0.1000, 0.6000, 0.8000, 0.1000, 0.6000]), tensor([0, 1, 0, 0, 1]))

    Notes:
        - `Metrics` stores the complete prediction and target history, which is memory-intensive
          but necessary for metrics like AUROC that operate on the entire dataset.
        - For metrics that can be meaningfully averaged batch-by-batch (like accuracy),
          consider using [`MetricMeter`][danling.metric.metric_meter.MetricMeter] for better memory efficiency.
        - The `ignore_index` parameter is useful for handling masked or padded values in classification tasks.
        - The `ignore_nan` parameter helps handle missing values in regression tasks.
        - All metrics are synchronized across devices in distributed training environments.

    See Also:
        - [`MetricMeters`][danling.metric.metric_meter.MetricMeters]:
            Memory-efficient metric tracker that averages multiple metrics batch-by-batch.
    """

    metrics: FlatDict[str, Callable]
    preprocess: Optional[Callable] = None
    ignore_index: int = -100
    ignore_nan: bool = True
    _input: Tensor
    _target: Tensor
    _inputs: Tensor
    _targets: Tensor

    def __init__(
        self,
        *args,
        device: torch.device | None = None,
        ignore_index: int | None = None,
        ignore_nan: bool | None = None,
        preprocess: Callable | None = None,
        **metrics: Callable,
    ):
        super().__init__(device=device)
        self.metrics = FlatDict()
        self._add_state("_input", torch.empty(0))
        self._add_state("_target", torch.empty(0))
        self._add_state("_inputs", torch.empty(0))
        self._add_state("_targets", torch.empty(0))
        self.world_size = get_world_size()
        if args:
            from .metric_meter import MetricMeters

            if len(args) == 1 and isinstance(args[0], MetricMeters):
                meters = args[0]
                for name, meter in meters.items():
                    metrics.setdefault(name, meter.metric)
                if preprocess is None:
                    preprocess = meters.getattr("preprocess")
                if ignore_index is None:
                    ignore_index = meters.ignore_index
                if ignore_nan is None:
                    ignore_nan = meters.ignore_nan
            else:
                for metric in args:
                    if not callable(metric):
                        raise ValueError(f"Expected metric to be callable, but got {type(metric)}")
                    metrics.setdefault(metric.__name__, metric)
        self.preprocess = preprocess
        if ignore_index is not None:
            self.ignore_index = ignore_index
        if ignore_nan is not None:
            self.ignore_nan = ignore_nan
        for name, metric in metrics.items():
            self.metrics[name] = self._preprocess_callable(metric)

    def update(self, input: Tensor | NestedTensor | Sequence, target: Tensor | NestedTensor | Sequence) -> None:
        # convert input and target to Tensor if they are not
        if not isinstance(input, (Tensor, NestedTensor)):
            try:
                input = torch.tensor(input)
            except ValueError:
                input = NestedTensor(input)
        if not isinstance(target, (Tensor, NestedTensor)):
            try:
                target = torch.tensor(target)
            except ValueError:
                target = NestedTensor(target)
        if input.ndim == target.ndim + 1:
            input = input.squeeze(-1)
        # convert input and target to NestedTensor if one of them is
        if isinstance(input, NestedTensor) or isinstance(target, NestedTensor):
            if isinstance(target, NestedTensor) and isinstance(input, NestedTensor):
                input, target = input.concat, target.concat
            elif isinstance(input, NestedTensor):
                input, mask = input.concat, input.mask
                target = target[mask]
            elif isinstance(target, NestedTensor):
                target, mask = target.concat, target.mask
                input = input[mask]
            else:
                raise ValueError(f"Unknown input and target: {input}, {target}")
        if self.preprocess is not None:
            input, target = self.preprocess(input, target, ignore_index=self.ignore_index, ignore_nan=self.ignore_nan)
        if self.world_size > 1:
            input, target = self._sync(input), self._sync(target)
        input, target = input.detach().to(self.device), target.detach().to(self.device)
        self._input = input
        self._target = target
        self._inputs = torch.cat([self._inputs, input]).to(input.dtype)
        self._targets = torch.cat([self._targets, target]).to(target.dtype)

    def value(self) -> NestedDict[str, float | flist]:
        return self.calculate(self.input, self.target)

    def average(self) -> NestedDict[str, float | flist]:
        return self.calculate(self.inputs, self.targets)

    def compute(self) -> NestedDict[str, float | flist]:
        return self.average()

    @property
    def val(self) -> NestedDict[str, float | flist]:
        return self.value()

    @property
    def avg(self) -> NestedDict[str, float | flist]:
        return self.average()

    @torch.inference_mode()
    def calculate(self, input: Tensor, target: Tensor) -> NestedDict[str, flist | float]:
        if (
            isinstance(input, (Tensor, NestedTensor))
            and input.numel() == 0 == target.numel()
            or isinstance(input, (list, dict))
            and len(input) == 0 == len(target)
        ):
            return NestedDict({name: nan for name in self.metrics.keys()})
        ret = NestedDict()
        for name, metric in self.metrics.items():
            score = self._calculate(metric, input, target)
            if isinstance(score, Mapping):
                ret.merge(score)
            else:
                ret[name] = score
        return ret

    @torch.inference_mode()
    def _calculate(self, metric, input: Tensor, target: Tensor) -> flist | float:
        score = metric(input, target)
        if isinstance(score, Tensor):
            return score.item() if score.numel() == 1 else flist(score.tolist())
        return score

    @torch.inference_mode()
    def merge_state(self, metrics: Iterable):
        raise NotImplementedError()

    def _sync(self, tensor: Tensor):
        local_size = torch.tensor([tensor.shape[0]], dtype=torch.int64, device=tensor.device)
        size_list = [torch.zeros_like(local_size) for _ in range(self.world_size)]
        dist.all_gather(size_list, local_size)
        sizes = torch.cat(size_list)
        max_size = sizes.max()

        padded_tensor = torch.empty((max_size, *tensor.shape[1:]), dtype=tensor.dtype, device=tensor.device)
        padded_tensor[: tensor.shape[0]] = tensor
        gathered_tensors = [torch.empty_like(padded_tensor) for _ in range(self.world_size)]
        dist.all_gather(gathered_tensors, padded_tensor)
        slices = [gathered_tensors[i][: sizes[i]] for i in range(self.world_size) if sizes[i] > 0]
        return torch.cat(slices, dim=0)

    @property
    def input(self) -> Tensor:
        return self._input

    @property
    def target(self) -> Tensor:
        return self._target

    @property
    def inputs(self) -> Tensor:
        return self._inputs

    @property
    def targets(self) -> Tensor:
        return self._targets

    def __repr__(self):
        keys = tuple(i for i in self.metrics.keys())
        return f"{self.__class__.__name__}{keys}"

    def __format__(self, format_spec):
        val, avg = self.value(), self.average()
        return "\t".join(
            [f"{key}: {val[key].__format__(format_spec)} ({avg[key].__format__(format_spec)})" for key in self.metrics]
        )

    def reset(self: Self) -> Self:  # pragma: no cover
        r"""
        Reset the metric state variables to their default value.
        The tensors in the default values are also moved to the device of
        the last ``self.to(device)`` call.
        """
        for state_name, default in self._state_name_to_default.items():
            if isinstance(default, Tensor):
                setattr(self, state_name, default.clone().to(self.device))
            elif isinstance(default, list):
                setattr(
                    self,
                    state_name,
                    flist(tensor.clone().to(self.device) for tensor in default),
                )
            elif isinstance(default, dict):
                setattr(
                    self,
                    state_name,
                    DefaultDict(
                        lambda: torch.tensor(0.0, device=self.device),
                        {key: tensor.clone().to(self.device) for key, tensor in default.items()},
                    ),
                )
            elif isinstance(default, (int, float)):
                setattr(self, state_name, default)
            else:
                raise TypeError(
                    f"Invalid type for default value for {state_name}. Received {type(default)},"
                    "but expected ``Tensor``, a list of ``Tensor``,"
                    "a dictionary with ``Tensor``, int, or float."
                )
        return self

    def _preprocess_callable(self, func: Callable) -> Callable:
        if not callable(func):
            raise ValueError(f"Expected func to be callable, but got {type(func)}")
        if "preprocess" not in signature(func).parameters:
            return func
        return partial(func, preprocess=self.preprocess is None)

reset

Python
reset() -> Self

Reset the metric state variables to their default value. The tensors in the default values are also moved to the device of the last self.to(device) call.

Source code in danling/metric/metrics.py
Python
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
def reset(self: Self) -> Self:  # pragma: no cover
    r"""
    Reset the metric state variables to their default value.
    The tensors in the default values are also moved to the device of
    the last ``self.to(device)`` call.
    """
    for state_name, default in self._state_name_to_default.items():
        if isinstance(default, Tensor):
            setattr(self, state_name, default.clone().to(self.device))
        elif isinstance(default, list):
            setattr(
                self,
                state_name,
                flist(tensor.clone().to(self.device) for tensor in default),
            )
        elif isinstance(default, dict):
            setattr(
                self,
                state_name,
                DefaultDict(
                    lambda: torch.tensor(0.0, device=self.device),
                    {key: tensor.clone().to(self.device) for key, tensor in default.items()},
                ),
            )
        elif isinstance(default, (int, float)):
            setattr(self, state_name, default)
        else:
            raise TypeError(
                f"Invalid type for default value for {state_name}. Received {type(default)},"
                "but expected ``Tensor``, a list of ``Tensor``,"
                "a dictionary with ``Tensor``, int, or float."
            )
    return self

MultiTaskMetrics

Bases: MultiTaskDict

A container for managing multiple Metrics for multiple tasks.

Typically, we have multiple tasks, and each task has multiple metrics. For example, a multi-task model might have a classification task and a regression task. We want to compute auroc and accuracy for the classification task, and pearson and rmse for the regression task.

MultiTaskMetrics is a mapping from task names to Metrics instances.

Examples:

Python Console Session
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
>>> from danling.metric.functional import auroc, auprc, pearson, spearman, accuracy, mcc
>>> metrics = MultiTaskMetrics()
>>> metrics.dataset1.cls = Metrics(auroc=auroc, auprc=auprc)
>>> metrics.dataset1.reg = Metrics(pearson=pearson, spearman=spearman)
>>> metrics.dataset2 = Metrics(auroc=auroc, auprc=auprc)
>>> metrics
MultiTaskMetrics(...
  ('dataset1'): MultiTaskMetrics(...
    ('cls'): Metrics('auroc', 'auprc')
    ('reg'): Metrics('pearson', 'spearman')
  )
  ('dataset2'): Metrics('auroc', 'auprc')
)
>>> metrics.update({"dataset1.cls": {"input": [0.2, 0.4, 0.5, 0.7], "target": [0, 1, 0, 1]}, "dataset1.reg": {"input": [0.1, 0.4, 0.6, 0.8], "target": [0.2, 0.3, 0.5, 0.7]}, "dataset2": {"input": [0.1, 0.4, 0.6, 0.8], "target": [0, 1, 0, 1]}})
>>> f"{metrics:.4f}"
'dataset1.cls: auroc: 0.7500 (0.7500)\tauprc: 0.8333 (0.8333)\ndataset1.reg: pearson: 0.9691 (0.9691)\tspearman: 1.0000 (1.0000)\ndataset2: auroc: 0.7500 (0.7500)\tauprc: 0.8333 (0.8333)'
>>> metrics.setattr("return_average", True)
>>> metrics.update({"dataset1.cls": {"input": [0.1, 0.4, 0.6, 0.8], "target": [0, 0, 1, 0]}, "dataset1.reg": {"input": [0.2, 0.3, 0.5, 0.7], "target": [0.2, 0.4, 0.6, 0.8]}, "dataset2": {"input": [0.2, 0.3, 0.5, 0.7], "target": [0, 0, 1, 0]}})
>>> f"{metrics:.4f}"
'dataset1.cls: auroc: 0.6667 (0.7000)\tauprc: 0.5000 (0.5556)\ndataset1.reg: pearson: 0.9898 (0.9146)\tspearman: 1.0000 (0.9222)\ndataset2: auroc: 0.6667 (0.7333)\tauprc: 0.5000 (0.7000)'
>>> metrics.update({"dataset1": {"cls": {"input": [0.1, 0.4, 0.6, 0.8], "target": [1, 0, 1, 0]}}})
>>> f"{metrics:.4f}"
'dataset1.cls: auroc: 0.2500 (0.5286)\tauprc: 0.5000 (0.4789)\ndataset1.reg: pearson: 0.9898 (0.9146)\tspearman: 1.0000 (0.9222)\ndataset2: auroc: 0.6667 (0.7333)\tauprc: 0.5000 (0.7000)'
>>> metrics.update(dict(loss=""))
Traceback (most recent call last):
ValueError: Metric loss not found in ...
Notes
  • MultiTaskMetrics is a container for managing multiple Metrics instances
  • Each task has its own Metrics instance for tracking task-specific metrics
  • Updates are passed as a dictionary mapping task names to (input, target) pairs
  • Access values using attribute notation: metrics.task_name.metric_name
See Also
  • MultiTaskMetricMeters: Memory-efficient metric tracker that averages multiple metrics batch-by-batch for multi-task learning.
Source code in danling/metric/metrics.py
Python
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
class MultiTaskMetrics(MultiTaskDict):
    r"""
    A container for managing multiple `Metrics` for multiple tasks.

    Typically, we have multiple tasks, and each task has multiple metrics.
    For example, a multi-task model might have a classification task and a regression task.
    We want to compute auroc and accuracy for the classification task,
    and pearson and rmse for the regression task.

    `MultiTaskMetrics` is a mapping from task names to `Metrics` instances.

    Examples:
        >>> from danling.metric.functional import auroc, auprc, pearson, spearman, accuracy, mcc
        >>> metrics = MultiTaskMetrics()
        >>> metrics.dataset1.cls = Metrics(auroc=auroc, auprc=auprc)
        >>> metrics.dataset1.reg = Metrics(pearson=pearson, spearman=spearman)
        >>> metrics.dataset2 = Metrics(auroc=auroc, auprc=auprc)
        >>> metrics  # doctest: +ELLIPSIS
        MultiTaskMetrics(...
          ('dataset1'): MultiTaskMetrics(...
            ('cls'): Metrics('auroc', 'auprc')
            ('reg'): Metrics('pearson', 'spearman')
          )
          ('dataset2'): Metrics('auroc', 'auprc')
        )
        >>> metrics.update({"dataset1.cls": {"input": [0.2, 0.4, 0.5, 0.7], "target": [0, 1, 0, 1]}, "dataset1.reg": {"input": [0.1, 0.4, 0.6, 0.8], "target": [0.2, 0.3, 0.5, 0.7]}, "dataset2": {"input": [0.1, 0.4, 0.6, 0.8], "target": [0, 1, 0, 1]}})
        >>> f"{metrics:.4f}"
        'dataset1.cls: auroc: 0.7500 (0.7500)\tauprc: 0.8333 (0.8333)\ndataset1.reg: pearson: 0.9691 (0.9691)\tspearman: 1.0000 (1.0000)\ndataset2: auroc: 0.7500 (0.7500)\tauprc: 0.8333 (0.8333)'
        >>> metrics.setattr("return_average", True)
        >>> metrics.update({"dataset1.cls": {"input": [0.1, 0.4, 0.6, 0.8], "target": [0, 0, 1, 0]}, "dataset1.reg": {"input": [0.2, 0.3, 0.5, 0.7], "target": [0.2, 0.4, 0.6, 0.8]}, "dataset2": {"input": [0.2, 0.3, 0.5, 0.7], "target": [0, 0, 1, 0]}})
        >>> f"{metrics:.4f}"
        'dataset1.cls: auroc: 0.6667 (0.7000)\tauprc: 0.5000 (0.5556)\ndataset1.reg: pearson: 0.9898 (0.9146)\tspearman: 1.0000 (0.9222)\ndataset2: auroc: 0.6667 (0.7333)\tauprc: 0.5000 (0.7000)'
        >>> metrics.update({"dataset1": {"cls": {"input": [0.1, 0.4, 0.6, 0.8], "target": [1, 0, 1, 0]}}})
        >>> f"{metrics:.4f}"
        'dataset1.cls: auroc: 0.2500 (0.5286)\tauprc: 0.5000 (0.4789)\ndataset1.reg: pearson: 0.9898 (0.9146)\tspearman: 1.0000 (0.9222)\ndataset2: auroc: 0.6667 (0.7333)\tauprc: 0.5000 (0.7000)'
        >>> metrics.update(dict(loss=""))  # doctest: +ELLIPSIS
        Traceback (most recent call last):
        ValueError: Metric loss not found in ...

    Notes:
        - `MultiTaskMetrics` is a container for managing multiple `Metrics` instances
        - Each task has its own `Metrics` instance for tracking task-specific metrics
        - Updates are passed as a dictionary mapping task names to (input, target) pairs
        - Access values using attribute notation: `metrics.task_name.metric_name`

    See Also:
        - [`MultiTaskMetricMeters`][danling.metric.metric_meter.MultiTaskMetricMeters]:
            Memory-efficient metric tracker that averages multiple metrics batch-by-batch for multi-task learning.
    """  # noqa: E501

    def __init__(self, *args, **kwargs):
        super().__init__(*args, default_factory=MultiTaskMetrics, **kwargs)

    def update(
        self,
        values: Mapping[str, Mapping[str, Tensor | NestedTensor | Sequence]],
    ) -> None:
        r"""
        Updates the average and current value in all metrics.

        Args:
            values: Dict of values to be added to the average.

        Raises:
            ValueError: If the value is not an instance of (Mapping).
        """

        for metric, value in values.items():
            if metric not in self:
                raise ValueError(f"Metric {metric} not found in {self}")
            if isinstance(self[metric], MultiTaskMetrics):
                for name, met in self[metric].items():
                    if name in value:
                        val = value[name]
                        if isinstance(value, Mapping):
                            met.update(**val)
                        elif isinstance(value, Sequence):
                            met.update(*val)
                        else:
                            raise ValueError(f"Expected value to be a Mapping or Sequence, but got {type(value)}")
            elif isinstance(self[metric], (Metrics, Metric)):
                if isinstance(value, Mapping):
                    self[metric].update(**value)
                elif isinstance(value, Sequence):
                    self[metric].update(*value)
                else:
                    raise ValueError(f"Expected value to be a Mapping or Sequence, but got {type(value)}")
            else:
                raise ValueError(
                    f"Expected {metric} to be an instance of MultiTaskMetrics, Metrics, or Metric, "
                    "but got {type(self[metric])}"
                )

    def set(  # pylint: disable=W0237
        self,
        name: str,
        metric: Metrics | Metric,  # type: ignore[override]
    ) -> None:
        if not isinstance(metric, (Metrics, Metric)):
            raise ValueError(f"Expected {metric} to be an instance of Metrics or Metric, but got {type(metric)}")
        super().set(name, metric)

update

Python
update(
    values: Mapping[
        str, Mapping[str, Tensor | NestedTensor | Sequence]
    ]
) -> None

Updates the average and current value in all metrics.

Parameters:

Name Type Description Default
values
Mapping[str, Mapping[str, Tensor | NestedTensor | Sequence]]

Dict of values to be added to the average.

 required

Raises:

Type Description
ValueError

If the value is not an instance of (Mapping).
Source code in danling/metric/metrics.py
Python
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
def update(
    self,
    values: Mapping[str, Mapping[str, Tensor | NestedTensor | Sequence]],
) -> None:
    r"""
    Updates the average and current value in all metrics.

    Args:
        values: Dict of values to be added to the average.

    Raises:
        ValueError: If the value is not an instance of (Mapping).
    """

    for metric, value in values.items():
        if metric not in self:
            raise ValueError(f"Metric {metric} not found in {self}")
        if isinstance(self[metric], MultiTaskMetrics):
            for name, met in self[metric].items():
                if name in value:
                    val = value[name]
                    if isinstance(value, Mapping):
                        met.update(**val)
                    elif isinstance(value, Sequence):
                        met.update(*val)
                    else:
                        raise ValueError(f"Expected value to be a Mapping or Sequence, but got {type(value)}")
        elif isinstance(self[metric], (Metrics, Metric)):
            if isinstance(value, Mapping):
                self[metric].update(**value)
            elif isinstance(value, Sequence):
                self[metric].update(*value)
            else:
                raise ValueError(f"Expected value to be a Mapping or Sequence, but got {type(value)}")
        else:
            raise ValueError(
                f"Expected {metric} to be an instance of MultiTaskMetrics, Metrics, or Metric, "
                "but got {type(self[metric])}"
            )

catch

Python
catch(
    error: Exceptions = Exception,
    exclude: Exceptions | None = None,
    callback: Callable = print_exc,
    *callback_args,
    **callback_kwargs
)

Decorator to catch error except for exclude. Detailed traceback will be printed to stderr.

catch is extremely useful for unfatal errors. For example, Runner saves checkpoint regularly, however, this might break running if the space is full. Decorating save method with catch will allow you to catch these errors and continue your running.

Parameters:

Name Type Description Default

error

Exceptions

Exceptions to be caught.

 Exception

exclude

Exceptions | None

Exceptions to be excluded.

 None

callback

Callable

Callback to be called when an error occurs. The first four arguments to callback are exc, func, args, kwargs. Additional arguments should be passed with *callback_args and **callback_kwargs.

 print_exc

callback_args

Arguments to be passed to callback.

 ()

callback_kwargs

Keyword arguments to be passed to callback.

 {}

Examples:

Python Console Session
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
>>> def file_not_found(*args, **kwargs):
...     raise FileNotFoundError
>>> func = file_not_found
>>> func()
Traceback (most recent call last):
FileNotFoundError
>>> func = catch(OSError)(file_not_found)
>>> func()
>>> func = catch(IOError)(file_not_found)
>>> func()
>>> func = catch(ZeroDivisionError)(file_not_found)
>>> func()
Traceback (most recent call last):
FileNotFoundError
Source code in danling/utils/decorators.py
Python
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
@flexible_decorator
def catch(  # pylint: disable=keyword-arg-before-vararg
    error: Exceptions = Exception,
    exclude: Exceptions | None = None,
    callback: Callable = print_exc,
    *callback_args,
    **callback_kwargs,
):
    r"""
    Decorator to catch `error` except for `exclude`.
    Detailed traceback will be printed to `stderr`.

    `catch` is extremely useful for unfatal errors.
    For example, `Runner` saves checkpoint regularly, however, this might break running if the space is full.
    Decorating `save` method with `catch` will allow you to catch these errors and continue your running.

    Args:
        error: Exceptions to be caught.
        exclude: Exceptions to be excluded.
        callback: Callback to be called when an error occurs.
            The first four arguments to `callback` are `exc`, `func`, `args`, `kwargs`.
            Additional arguments should be passed with `*callback_args` and `**callback_kwargs`.
        callback_args: Arguments to be passed to `callback`.
        callback_kwargs: Keyword arguments to be passed to `callback`.

    Examples:
        >>> def file_not_found(*args, **kwargs):
        ...     raise FileNotFoundError
        >>> func = file_not_found
        >>> func()
        Traceback (most recent call last):
        FileNotFoundError
        >>> func = catch(OSError)(file_not_found)
        >>> func()
        >>> func = catch(IOError)(file_not_found)
        >>> func()
        >>> func = catch(ZeroDivisionError)(file_not_found)
        >>> func()
        Traceback (most recent call last):
        FileNotFoundError
    """

    def decorator(func):
        @wraps(func)
        def wrapper(*args, **kwargs):  # pylint: disable=inconsistent-return-statements
            try:
                return func(*args, **kwargs)
            except error as exc:  # pylint: disable=broad-exception-caught
                if exclude is not None and isinstance(exc, exclude):
                    raise exc
                callback(exc, func, args, kwargs, *callback_args, **callback_kwargs)

        return wrapper

    decorator.__doc__ = catch.__doc__

    return decorator

debug

Python
debug(
    enable: bool = True,
    error: Exceptions = Exception,
    exclude: Optional[Exceptions] = None,
)

Contextmanager to enter debug mode on error except for exclude.

debug is intended to be used to catch the error and enter debug mode. Since it is mainly for development purposed, we include an enable args so that it can be deactivated.

Parameters:

Name Type Description Default

enable

bool

Whether to enable the contextmanager. Defaults to True.

 True

error

Exceptions

The error to catch. Defaults to Exception.

 Exception

exclude

Optional[Exceptions]

The error to exclude. Defaults to None.

 None
Source code in danling/utils/contextmanagers.py
Python
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
@contextmanager
def debug(
    enable: bool = True,
    error: Exceptions = Exception,
    exclude: Optional[Exceptions] = None,
):
    """
    Contextmanager to enter debug mode on `error` except for `exclude`.

    `debug` is intended to be used to catch the error and enter debug mode.
    Since it is mainly for development purposed, we include an `enable` args so that it can be deactivated.

    Args:
        enable: Whether to enable the contextmanager.
            Defaults to `True`.
        error: The error to catch.
            Defaults to `Exception`.
        exclude: The error to exclude.
            Defaults to `None`.
    """

    if not enable:
        yield
        return
    try:
        yield
    except error as exc:  # pylint: disable=broad-exception-caught
        if exclude is not None and isinstance(exc, exclude):
            raise exc
        _, m, tb = sys.exc_info()
        print(repr(m), file=sys.stderr)
        pdb.post_mortem(tb)
    finally:
        pass

flexible_decorator

Python
flexible_decorator(
    maybe_decorator: Optional[Callable] = None,
)

Meta decorator to allow bracket-less decorator when no arguments are passed.

Examples:

For decorator defined as follows:

Python Console Session
1
2
3
4
5
>>> @flexible_decorator
... def decorator(*args, **kwargs):
...     def wrapper(func, *args, **kwargs):
...         pass
...     return wrapper

The following two are equivalent:

Python Console Session
1
2
3
>>> @decorator
... def func(*args, **kwargs):
...     pass
Python Console Session
1
2
3
>>> @decorator()
... def func(*args, **kwargs):
...     pass
Source code in danling/utils/decorators.py
Python
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
def flexible_decorator(maybe_decorator: Optional[Callable] = None):
    r"""
    Meta decorator to allow bracket-less decorator when no arguments are passed.

    Examples:
        For decorator defined as follows:

        >>> @flexible_decorator
        ... def decorator(*args, **kwargs):
        ...     def wrapper(func, *args, **kwargs):
        ...         pass
        ...     return wrapper

        The following two are equivalent:

        >>> @decorator
        ... def func(*args, **kwargs):
        ...     pass

        >>> @decorator()
        ... def func(*args, **kwargs):
        ...     pass
    """

    def decorator(func: Callable):
        @wraps(func)
        def wrapper(*args, **kwargs):
            if len(args) == 1 and isfunction(args[0]):
                return func(**kwargs)(args[0])
            return func(*args, **kwargs)

        return wrapper

    if maybe_decorator is None:
        return decorator
    return decorator(maybe_decorator)

is_json_serializable

Python
is_json_serializable(obj: Any) -> bool

Check if obj is JSON serializable.

Source code in danling/utils/io.py
Python
181
182
183
184
185
186
187
188
189
def is_json_serializable(obj: Any) -> bool:
    r"""
    Check if `obj` is JSON serializable.
    """
    try:
        json.dumps(obj)
        return True
    except (TypeError, OverflowError):
        return False

load

Python
load(
    file: PathStr,
    *args: List[Any],
    **kwargs: Dict[str, Any]
) -> Any

Load any file with supported extensions.

Source code in danling/utils/io.py
Python
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
def load(file: PathStr, *args: List[Any], **kwargs: Dict[str, Any]) -> Any:
    r"""
    Load any file with supported extensions.
    """
    if not os.path.isfile(file):
        raise ValueError(f"Trying to load {file!r} but it is not a file.")
    extension = os.path.splitext(file)[-1].lower()[1:]
    if extension in PYTORCH:
        if not TORCH_AVAILABLE:
            raise ImportError(f"Trying to load {file!r} but torch is not installed.")
        return torch.load(file, *args, **kwargs)
    if extension in NUMPY:
        if not NUMPY_AVAILABLE:
            raise ImportError(f"Trying to load {file!r} but numpy is not installed.")
        return numpy.load(file, *args, **kwargs)
    if extension in JSON:
        with open(file) as fp:
            return json.load(fp, *args, **kwargs)  # type: ignore[arg-type]
    if extension in YAML:
        with open(file) as fp:
            kwargs.setdefault("Loader", yaml.FullLoader)  # type: ignore[arg-type]
            return yaml.load(fp, *args, **kwargs)  # type: ignore[arg-type]
    if extension in PICKLE:
        with open(file, "rb") as fp:
            return pickle.load(fp, *args, **kwargs)  # type: ignore[arg-type]
    if extension in PANDAS_SUPPORTED:
        return load_pandas(file, *args, **kwargs)
    raise ValueError(f"Tying to load {file!r} with unsupported extension={extension!r}")

load_pandas

Python
load_pandas(
    file: PathStr,
    *args: List[Any],
    **kwargs: Dict[str, Any]
) -> Any

Load any pandas data file with supported extensions.

Source code in danling/utils/io.py
Python
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
def load_pandas(file: PathStr, *args: List[Any], **kwargs: Dict[str, Any]) -> Any:
    r"""
    Load any pandas data file with supported extensions.
    """
    if not PANDAS_AVAILABLE:
        raise ImportError(f"Trying to load {file!r} but pandas is not installed.")
    if not os.path.isfile(file):
        raise ValueError(f"Trying to load {file!r} but it is not a file.")
    extension = os.path.splitext(file)[-1].lower()[1:]
    if extension in PANDAS or extension in PICKLE:
        return pandas.read_pickle(file, *args, **kwargs)
    if extension in PARQUET:
        return pandas.read_parquet(file, *args, **kwargs)
    if extension in H5:
        return pandas.read_hdf(file, *args, **kwargs)
    if extension in CSV:
        return pandas.read_csv(file, *args, **kwargs)
    if extension in JSON:
        return pandas.read_json(file, *args, **kwargs)
    if extension in EXCEL:
        return pandas.read_excel(file, *args, **kwargs)
    if extension in XML:
        return pandas.read_xml(file, *args, **kwargs)
    if extension in SQL:
        return pandas.read_sql(file, *args, **kwargs)
    raise ValueError(f"Tying to load {file!r} with unsupported extension={extension!r}")

method_cache

Python
method_cache(
    maxsize: int | None = 128, typed: bool = False
)

Decorator to cache the result of an instance method.

functools.lru_cache uses a strong reference to the instance, which will make the instance immortal and break the garbage collection.

method_cache uses a weak reference to the instance to resolve this issue.

See Also
Source code in danling/utils/decorators.py
Python
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
@flexible_decorator
def method_cache(maxsize: int | None = 128, typed: bool = False):
    r"""
    Decorator to cache the result of an instance method.

    `functools.lru_cache` uses a strong reference to the instance,
    which will make the instance immortal and break the garbage collection.

    `method_cache` uses a weak reference to the instance to resolve this issue.

    See Also:
        https://rednafi.github.io/reflections/dont-wrap-instance-methods-with-functoolslru_cache-decorator-in-python.html
    """

    def decorator(func):
        @wraps(func)
        def wrapper(self, *args, **kwargs):
            self_ref = ref(self)

            @wraps(func)
            @lru_cache(maxsize=maxsize, typed=typed)
            def cached_method(*args, **kwargs):
                return func(self_ref(), *args, **kwargs)

            setattr(self, func.__name__, cached_method)
            return cached_method(*args, **kwargs)

        return wrapper

    return decorator

save

Python
save(
    obj: Any,
    file: PathStr,
    *args: List[Any],
    **kwargs: Dict[str, Any]
) -> File

Save any file with supported extensions.

Source code in danling/utils/io.py
Python
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
def save(obj: Any, file: PathStr, *args: List[Any], **kwargs: Dict[str, Any]) -> File:
    r"""
    Save any file with supported extensions.
    """
    extension = os.path.splitext(file)[-1].lower()[1:]
    if extension in PYTORCH:
        if not TORCH_AVAILABLE:
            raise ImportError(f"Trying to save {obj} to {file!r} but torch is not installed.")
        torch.save(obj, file, *args, **kwargs)
    elif extension in NUMPY:
        if not NUMPY_AVAILABLE:
            raise ImportError(f"Trying to save {obj} to {file!r} but numpy is not installed.")
        numpy.save(file, obj, *args, **kwargs)
    elif extension in PANDAS:
        if not PANDAS_AVAILABLE:
            raise ImportError(f"Trying to save {obj} to {file!r} but pandas is not installed.")
        pandas.to_pickle(obj, file, *args, **kwargs)
    elif extension in PARQUET:
        if isinstance(obj, pandas.DataFrame):
            obj.to_parquet(file, *args, **kwargs)
        elif not PYARROW_AVAILABLE:
            raise ImportError(f"Trying to save {obj} to {file!r} but pyarrow is not installed.")
        else:
            pyarrow.parquet.write_table(obj, file, *args, **kwargs)
    elif extension in CSV:
        if isinstance(obj, pandas.DataFrame):
            obj.to_csv(file, *args, **kwargs)
        else:
            raise NotImplementedError(f"Trying to save {obj} to {file!r} but is not supported")
    elif extension in JSON:
        if isinstance(obj, FlatDict):
            obj.json(file)
        else:
            with open(file, "w") as fp:
                json.dump(obj, fp, *args, **kwargs)  # type: ignore[arg-type]
    elif extension in YAML:
        if isinstance(obj, FlatDict):
            obj.yaml(file)
        else:
            with open(file, "w") as fp:
                yaml.dump(obj, fp, *args, **kwargs)  # type: ignore[arg-type, call-overload]
    elif extension in PICKLE:
        with open(file, "wb") as fp:
            pickle.dump(obj, fp, *args, **kwargs)  # type: ignore[arg-type]
    else:
        raise ValueError(f"Tying to save {obj} to {file!r} with unsupported extension={extension!r}")
    return file