Skip to content

DanLing

danling

AverageMeter

Computes and stores the average and current value.

Attributes:

Name Type Description
val float

Results of current batch on current device.
bat

Results of current batch on all devices.
avg

Results of all results on all devices.
sum float

Sum of values.
count float

Number of values.
See Also

[MetricMeter]: Average Meter with metric function built-in. [AverageMeters]: Manage multiple average meters in one object. [MultiTaskAverageMeters]: Manage multiple average meters in one object with multi-task support.

Examples:

Python Console Session
>>> meter = AverageMeter()
>>> meter.update(0.7)
>>> meter.val
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
>>> meter.reset()
>>> meter.val
0
>>> meter.avg
nan
Source code in danling/metrics/average_meter.py
Python
 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
class AverageMeter:
    r"""
    Computes and stores the average and current value.

    Attributes:
        val: Results of current batch on current device.
        bat: Results of current batch on all devices.
        avg: Results of all results on all devices.
        sum: Sum of values.
        count: Number of values.

    See Also:
        [`MetricMeter`]: Average Meter with metric function built-in.
        [`AverageMeters`]: Manage multiple average meters in one object.
        [`MultiTaskAverageMeters`]: Manage multiple average meters in one object with multi-task support.

    Examples:
        >>> meter = AverageMeter()
        >>> meter.update(0.7)
        >>> meter.val
        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
        >>> meter.reset()
        >>> meter.val
        0
        >>> meter.avg
        nan
    """

    val: float = 0
    n: float = 1
    sum: float = 0
    count: float = 0

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

    def reset(self) -> None:
        r"""
        Resets the meter.
        """

        self.val = 0
        self.n = 1
        self.sum = 0
        self.count = 0

    def update(self, value, n: float = 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.val = value
        self.n = n
        self.sum += value * n
        self.count += n

    def value(self):
        return self.val

    def batch(self):
        world_size = get_world_size()
        if world_size == 1:
            return self.val / self.n if self.n != 0 else float("nan")
        synced_tuple = [None for _ in range(world_size)]
        dist.all_gather_object(synced_tuple, (self.val * self.n, self.n))
        val, n = zip(*synced_tuple)
        count = sum(n)
        if count == 0:
            return float("nan")
        return sum(val) / count

    def average(self):
        world_size = get_world_size()
        if world_size == 1:
            return self.sum / self.count if self.count != 0 else float("nan")
        synced_tuple = [None for _ in range(world_size)]
        dist.all_gather_object(synced_tuple, (self.sum, self.count))
        val, n = zip(*synced_tuple)
        count = sum(n)
        if count == 0:
            return float("nan")
        return sum(val) / count

    @property
    def bat(self):
        return self.batch()

    @property
    def avg(self):
        return self.average()

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

reset

Python
1
reset() -> None

Resets the meter.

Source code in danling/metrics/average_meter.py
Python
76
77
78
79
80
81
82
83
84
def reset(self) -> None:
    r"""
    Resets the meter.
    """

    self.val = 0
    self.n = 1
    self.sum = 0
    self.count = 0

update

Python
1
update(value, n: float = 1) -> 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 float

Number of values to be added.

 1
Source code in danling/metrics/average_meter.py
Python
86
87
88
89
90
91
92
93
94
95
96
97
98
def update(self, value, n: float = 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.val = value
    self.n = n
    self.sum += value * n
    self.count += n

AverageMeters

Bases: MetricsDict

Manages multiple average meters in one object.

See Also

[AverageMeter]: Computes and stores the average and current value. [MultiTaskAverageMeters]: Manage multiple average meters in one object with multi-task support. [MetricMeters]: Manage multiple metric meters in one object.

Examples:

Python Console Session
>>> 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()
>>> f"{meters:.4f}"
'loss: 0.0000 (nan)\tauroc: 0.0000 (nan)\tr2: 0.0000 (nan)'
Source code in danling/metrics/average_meter.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
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
class AverageMeters(MetricsDict):
    r"""
    Manages multiple average meters in one object.

    See Also:
        [`AverageMeter`]: Computes and stores the average and current value.
        [`MultiTaskAverageMeters`]: Manage multiple average meters in one object with multi-task support.
        [`MetricMeters`]: Manage multiple metric 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()
        >>> f"{meters:.4f}"
        'loss: 0.0000 (nan)\tauroc: 0.0000 (nan)\tr2: 0.0000 (nan)'
    """

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

    @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 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
1
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/metrics/average_meter.py
Python
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
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 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

Computes metrics and averages them over time.

Attributes:

Name Type Description
metric Callable

Metric function for computing the value.
ignored_index Optional[int]

Index to be ignored in the computation.
val Optional[int]

Results of current batch on current device.
bat Optional[int]

Results of current batch on all devices.
avg Optional[int]

Results of all results on all devices.
sum Optional[int]

Sum of values.
count Optional[int]

Number of values.
See Also

[AverageMeter]: Average meter for computed values. [MetricMeters]: Manage multiple metric meters in one object.

Examples:

Python Console Session
>>> from danling.metrics.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()
>>> meter.val
0
>>> meter.avg
nan
Source code in danling/metrics/metric_meter.py
Python
 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
class MetricMeter(AverageMeter):
    r"""
    Computes metrics and averages them over time.

    Attributes:
        metric: Metric function for computing the value.
        ignored_index: Index to be ignored in the computation.
        val: Results of current batch on current device.
        bat: Results of current batch on all devices.
        avg: Results of all results on all devices.
        sum: Sum of values.
        count: Number of values.

    See Also:
        [`AverageMeter`]: Average meter for computed values.
        [`MetricMeters`]: Manage multiple metric meters in one object.

    Examples:
        >>> from danling.metrics.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()
        >>> meter.val
        0
        >>> meter.avg
        nan
    """

    metric: Callable
    ignored_index: Optional[int] = None

    def __init__(self, metric: Callable, ignored_index: int | None = None) -> None:
        self.metric = metric
        self.ignored_index = ignored_index
        super().__init__()

    def reset(self) -> None:
        r"""
        Resets the meter.
        """

        self.val = 0
        self.n = 1
        self.sum = 0
        self.count = 0

    def update(  # type: ignore[override] # pylint: disable=W0237
        self,
        input: Tensor | NestedTensor | Sequence,  # pylint: disable=W0622
        target: Tensor | NestedTensor | Sequence,
    ) -> 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.
        """

        input, target = preprocess(input, target, ignored_index=self.ignored_index)
        super().update(self.metric(input, target).item(), n=len(input))

reset

Python
1
reset() -> None

Resets the meter.

Source code in danling/metrics/metric_meter.py
Python
80
81
82
83
84
85
86
87
88
def reset(self) -> None:
    r"""
    Resets the meter.
    """

    self.val = 0
    self.n = 1
    self.sum = 0
    self.count = 0

update

Python
1
update(input: Tensor | NestedTensor | Sequence, target: Tensor | NestedTensor | Sequence) -> 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/metrics/metric_meter.py
Python
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
def update(  # type: ignore[override] # pylint: disable=W0237
    self,
    input: Tensor | NestedTensor | Sequence,  # pylint: disable=W0622
    target: Tensor | NestedTensor | Sequence,
) -> 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.
    """

    input, target = preprocess(input, target, ignored_index=self.ignored_index)
    super().update(self.metric(input, target).item(), n=len(input))

MetricMeters

Bases: AverageMeters

Manages multiple metric meters in one object.

Attributes:

Name Type Description
ignored_index Optional[int]

Index to be ignored in the computation. Defaults to None.
See Also

[MetricMeter]: Computes metrics and averages them over time. [AverageMeters]: Average meters for computed values.

from danling.metrics.functional import accuracy, auroc, auprc meters = MetricMeters(acc=accuracy, auroc=auroc, auprc=auprc) 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’

Source code in danling/metrics/metric_meter.py
Python
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
class MetricMeters(AverageMeters):
    r"""
    Manages multiple metric meters in one object.

    Attributes:
        ignored_index: Index to be ignored in the computation.
            Defaults to None.

    See Also:
        [`MetricMeter`]: Computes metrics and averages them over time.
        [`AverageMeters`]: Average meters for computed values.

    >>> from danling.metrics.functional import accuracy, auroc, auprc
    >>> meters = MetricMeters(acc=accuracy, auroc=auroc, auprc=auprc)
    >>> 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'
    """

    ignored_index: Optional[int] = None

    def __init__(self, *args, ignored_index: int | None = None, **kwargs) -> None:
        self.setattr("ignored_index", ignored_index)
        for meter in args:
            if callable(meter):
                meter = MetricMeter(meter, ignored_index=self.ignored_index)
            if not isinstance(meter, MetricMeter):
                raise ValueError(f"Expected meter to be an instance of MetricMeter, but got {type(meter)}")
        for name, meter in kwargs.items():
            if callable(meter):
                kwargs[name] = meter = MetricMeter(meter, ignored_index=self.ignored_index)
            if not isinstance(meter, MetricMeter):
                raise ValueError(f"Expected {name} to be an instance of MetricMeter, but got {type(meter)}")
        super().__init__(*args, default_factory=None, **kwargs)  # 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.
        """

        input, target = preprocess(input, target, ignored_index=self.ignored_index)
        for meter in self.values():
            meter.update(input, target)

    def set(self, name: str, meter: MetricMeter | Callable) -> None:  # type: ignore[override] # pylint: disable=W0237
        if callable(meter):
            meter = MetricMeter(meter, ignored_index=self.ignored_index)
        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 __repr__(self):
        keys = tuple(i for i in self.keys())
        return f"{self.__class__.__name__}{keys}"

update

Python
1
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/metrics/metric_meter.py
Python
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
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.
    """

    input, target = preprocess(input, target, ignored_index=self.ignored_index)
    for meter in self.values():
        meter.update(input, target)

MultiTaskAverageMeters

Bases: MultiTaskDict

Manages multiple average meters in one object with multi-task support.

See Also

[AverageMeter]: Computes and stores the average and current value. [AverageMeters]: Manage multiple average meters in one object. [MetricMeters]: Manage multiple metric meters in one object.

Examples:

Python Console Session
>>> 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()
>>> 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)'
Source code in danling/metrics/average_meter.py
Python
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
class MultiTaskAverageMeters(MultiTaskDict):
    r"""
    Manages multiple average meters in one object with multi-task support.

    See Also:
        [`AverageMeter`]: Computes and stores the average and current value.
        [`AverageMeters`]: Manage multiple average meters in one object.
        [`MetricMeters`]: Manage multiple metric meters in one object.

    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()
        >>> 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)'
    """  # 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)

update

Python
1
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/metrics/average_meter.py
Python
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
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
>>> from danling.metrics.functional import accuracy
>>> metrics = MultiTaskMetricMeters()
>>> metrics.dataset1.cls = MetricMeters(acc=accuracy)
>>> metrics.dataset2 = MetricMeters(acc=accuracy)
>>> metrics
MultiTaskMetricMeters(<class 'danling.metrics.metric_meter.MultiTaskMetricMeters'>,
  ('dataset1'): MultiTaskMetricMeters(<class 'danling.metrics.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 ...
Source code in danling/metrics/metric_meter.py
Python
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
class MultiTaskMetricMeters(MultiTaskAverageMeters):
    r"""
    Examples:
        >>> from danling.metrics.functional import accuracy
        >>> metrics = MultiTaskMetricMeters()
        >>> metrics.dataset1.cls = MetricMeters(acc=accuracy)
        >>> metrics.dataset2 = MetricMeters(acc=accuracy)
        >>> metrics
        MultiTaskMetricMeters(<class 'danling.metrics.metric_meter.MultiTaskMetricMeters'>,
          ('dataset1'): MultiTaskMetricMeters(<class 'danling.metrics.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 ...
    """  # 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,
        metric: MetricMeter | MetricMeters | Callable,  # type: ignore[override]
    ) -> None:
        if callable(metric):
            metric = MetricMeter(metric)
        if not isinstance(metric, (MetricMeter, MetricMeters)):
            raise ValueError(
                f"Expected {metric} to be an instance of MetricMeter or MetricMeters, but got {type(metric)}"
            )
        super().set(name, metric)

update

Python
1
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/metrics/metric_meter.py
Python
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
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
>>> 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
 25
 26
 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
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 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.

AccelerateRunner uses [accelerate][accelerate] as distributed backend to provide seamless distributed training experience.

AccelerateRunner will automatically prepare everything, including model, criterion, optimizer, scheduler, and dataloaders for distribute training, mixed precision, and deepspeed (optional).

In fact, you don’t even need to create dataloaders, just define datasets and AccelerateRunner will create dataloaders for you. AccelerateRunner will inspect the train flag in corresponding dataset to set shuffle and drop_last automatically.

Source code in danling/runner/accelerate_runner.py
Python
 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
class AccelerateRunner(TorchRunner, Accelerator):  # pylint: disable=too-many-public-methods
    r"""
    Set up everything for running a job.

    `AccelerateRunner` uses [`accelerate`][accelerate] as distributed backend to
    provide seamless distributed training experience.

    `AccelerateRunner` will automatically [`prepare`][accelerate.Accelerator.prepare] everything,
    including `model`, `criterion`, `optimizer`, `scheduler`, and `dataloaders` for distribute training,
    mixed precision, and deepspeed (optional).

    In fact, you don't even need to create `dataloaders`, just define
    `datasets` and `AccelerateRunner` will create `dataloaders` for you.
    `AccelerateRunner` will inspect the `train` flag in corresponding dataset to
    set `shuffle` and `drop_last` automatically.
    """

    _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(len(d) for d in self.dataloaders.values()) // 10
        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:
            zero_grad: Whether to zero the gradients.
        """

        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.get("max_grad_value"))
            if self.config.get("max_grad_norm") is not None:
                self.clip_grad_norm_(self.model.parameters(), self.config.get("max_grad_norm"))
        self.optimizer.step()
        if self.scheduler is not None:
            self.scheduler.step()
        self.optimizer.zero_grad()
        self.config.step = 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
1
advance(loss) -> None

Backward loss and step optimizer & scheduler.

Parameters:

Name Type Description Default
zero_grad

Whether to zero the gradients.

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

    Args:
        zero_grad: Whether to zero the gradients.
    """

    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.get("max_grad_value"))
        if self.config.get("max_grad_norm") is not None:
            self.clip_grad_norm_(self.model.parameters(), self.config.get("max_grad_norm"))
    self.optimizer.step()
    if self.scheduler is not None:
        self.scheduler.step()
    self.optimizer.zero_grad()
    self.config.step = self.step

BaseRunner

Base class for all runners.

BaseRunner sets up basic running environment, including seed, deterministic, and logging.

BaseRunner also provides some basic methods, such as, step, state_dict, save_checkpoint, load_checkpoint.

BaseRunner defines all basic attributes and relevant properties such as scores, progress, etc.

ID:

Name Type Description
timestamp str

A time string representing the creation time of run.
name str

f"{self.config.experiment_name}-{self.config.run_name}".
id str

f"{self.config.experiment_id:.8}{self.config.run_id:.8}".
uuid (UUID, property)

uuid5(self.config.run_id, self.id).

Core:

Name Type Description
mode (RunnerMode, property)

Running mode.
config Config

Running config. See [Config] for details.

Model:

Name Type Description
model Callable
criterion Callable
optimizer Any | None
scheduler Any | None

Data:

Name Type Description
datasets FlatDict

All datasets, should be in the form of {subset: dataset}. Initialised to FlatDict by default.
datasamplers FlatDict

All datasamplers, should be in the form of {subset: datasampler}. Initialised to FlatDict by default.
dataloaders FlatDict

All dataloaders, should be in the form of {subset: dataloader}. Initialised to FlatDict by default.
split str

Current running split.
batch_size (int, property)

Number of samples per batch in current running split.
batch_size_equivalent (int, property)

Total batch_size (batch_size * world_size * accum_steps).

datasets, datasamplers, dataloaders should be a dict with the same keys. Their keys should be split (e.g. train, val, test).

Progress:

Name Type Description
progress (float, property)

Running Progress, in range(0, 1).

Results:

Name Type Description
results NestedDict

Results include all metric information of the model. Results should be in the form of {epoch: {subset: {metric: score}}}.
latest_result (NestedDict, property)

Most recent result, should be in the form of {subset: {metric: score}}.
best_result (NestedDict, property)

Best result, should be in the form of {subset: {metric: score}}.
scores (List[float], property)

Score is the core metric that is used to evaluate the performance of the model. Scores should be in the form of {epoch: score}.
latest_score (float, property)

Most recent score, should be in the form of score.
best_score (float, property)

Best score, should be in the form of score.
score_split Optional[str]

The subset to calculate the score. If is None, will use the last set of the result.
score_name str

The metric name of score. Defaults to "loss".
is_best (bool, property)

If latest_score == best_score.

A result is a dict with the same split as keys, like dataloaders. A typical result shall look like this:

Python
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
{
    "train": {
        "loss": 0.1,
        "accuracy": 0.9,
    },
    "val": {
        "loss": 0.2,
        "accuracy": 0.8,
    },
    "test": {
        "loss": 0.3,
        "accuracy": 0.7,
    },
}

scores are dynamically extracted from results by score_split and score_name. They represent the core metric that is used in comparing the performance against different models and settings. For the above results, If score_split = "val", score_name = "accuracy", then scores = 0.9.

IO:

Name Type Description
dir (str, property)

Directory of the run. Defaults to ${self.project_root}/${self.name}-${self.id}/${self.timestamp}).
checkpoint_dir (str, property)

Directory of checkpoints.
log_path (str, property)

Path of log file.
checkpoint_dir_name str

The name of the directory under runner.dir to save checkpoints. Defaults to "checkpoints".

Parallel Training:

Name Type Description
world_size (int, property)

Number of processes.
rank (int, property)

Process index of all processes.
local_rank (int, property)

Process index of local processes.
distributed (bool, property)

If runner is running in distributed mode.
is_main_process (bool, property)

If current process is the main process of all processes.
is_local_main_process (bool, property)

If current process is the main process of local processes.

logging:

Name Type Description
meters AverageMeters | MultiTaskAverageMeters

Average meters. Initialised to AverageMeters by default.
metrics Metrics | MultiTaskMetrics | MetricMeters | None

Metrics for evaluating.
logger Logger | None
writer Any | None
See Also

Config: The runeer base that stores runtime information. BaseRunner: The base runner class.

Source code in danling/runner/base_runner.py
Python
  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
class BaseRunner(metaclass=RunnerMeta):  # pylint: disable=too-many-public-methods
    r"""
    Base class for all runners.

    `BaseRunner` sets up basic running environment, including `seed`, `deterministic`, and `logging`.

    `BaseRunner` also provides some basic methods, such as, `step`, `state_dict`, `save_checkpoint`, `load_checkpoint`.

    `BaseRunner` defines all basic attributes and relevant properties such as `scores`, `progress`, etc.

    Attributes: ID:
        timestamp (str): A time string representing the creation time of run.
        name (str): `f"{self.config.experiment_name}-{self.config.run_name}"`.
        id (str): `f"{self.config.experiment_id:.8}{self.config.run_id:.8}"`.
        uuid (UUID, property): `uuid5(self.config.run_id, self.id)`.

    Attributes: Core:
        mode (RunnerMode, property): Running mode.
        config (Config): Running config. See [`Config`] for details.

    Attributes: Model:
        model (Callable):
        criterion (Callable):
        optimizer:
        scheduler:

    Attributes: Data:
        datasets (FlatDict): All datasets, should be in the form of ``{subset: dataset}``.
            Initialised to `FlatDict` by default.
        datasamplers (FlatDict): All datasamplers, should be in the form of ``{subset: datasampler}``.
            Initialised to `FlatDict` by default.
        dataloaders (FlatDict): All dataloaders, should be in the form of ``{subset: dataloader}``.
            Initialised to `FlatDict` by default.
        split (str): Current running split.
        batch_size (int, property): Number of samples per batch in current running split.
        batch_size_equivalent (int, property): Total batch_size (`batch_size * world_size * accum_steps`).

    `datasets`, `datasamplers`, `dataloaders` should be a dict with the same keys.
    Their keys should be `split` (e.g. `train`, `val`, `test`).

    Attributes: Progress:
        progress (float, property): Running Progress, in `range(0, 1)`.

    Attributes: Results:
        results (NestedDict): Results include all metric information of the model.
            Results should be in the form of `{epoch: {subset: {metric: score}}}`.
        latest_result (NestedDict, property): Most recent result, should be in the form of `{subset: {metric: score}}`.
        best_result (NestedDict, property): Best result, should be in the form of `{subset: {metric: score}}`.
        scores (List[float], property): Score is the core metric that is used to evaluate the performance of the model.
            Scores should be in the form of `{epoch: score}`.
        latest_score (float, property): Most recent score, should be in the form of `score`.
        best_score (float, property): Best score, should be in the form of `score`.
        score_split (Optional[str]): The subset to calculate the score.
            If is `None`, will use the last set of the result.
        score_name (str): The metric name of score.
            Defaults to `"loss"`.
        is_best (bool, property): If `latest_score == best_score`.

    A `result` is a dict with the same `split` as keys, like `dataloaders`.
    A typical `result` shall look like this:
    ```python
    {
        "train": {
            "loss": 0.1,
            "accuracy": 0.9,
        },
        "val": {
            "loss": 0.2,
            "accuracy": 0.8,
        },
        "test": {
            "loss": 0.3,
            "accuracy": 0.7,
        },
    }
    ```

    `scores` are dynamically extracted from `results` by `score_split` and `score_name`.
    They represent the core metric that is used in comparing the performance against different models and settings.
    For the above `results`, If `score_split = "val"`, `score_name = "accuracy"`, then `scores = 0.9`.

    Attributes: IO:
        dir (str, property): Directory of the run.
            Defaults to `${self.project_root}/${self.name}-${self.id}/${self.timestamp})`.
        checkpoint_dir (str, property): Directory of checkpoints.
        log_path (str, property):  Path of log file.
        checkpoint_dir_name (str): The name of the directory under `runner.dir` to save checkpoints.
            Defaults to `"checkpoints"`.

    Attributes: Parallel Training:
        world_size (int, property): Number of processes.
        rank (int, property): Process index of all processes.
        local_rank (int, property): Process index of local processes.
        distributed (bool, property): If runner is running in distributed mode.
        is_main_process (bool, property): If current process is the main process of all processes.
        is_local_main_process (bool, property): If current process is the main process of local processes.

    Attributes: logging:
        meters (AverageMeters | MultiTaskAverageMeters): Average meters.
            Initialised to `AverageMeters` by default.
        metrics (Metrics | MultiTaskMetrics | MetricMeters | None): Metrics for evaluating.
        logger:
        writer:

    See Also:
        [`Config`][danling.runner.Config]: The runeer base that stores runtime information.
        [`BaseRunner`][danling.runner.BaseRunner]: The base runner class.
    """

    # 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
    criterion: Callable | None = None
    optimizer: Any | None = None
    scheduler: Any | None = None

    datasets: FlatDict
    datasamplers: FlatDict
    dataloaders: FlatDict
    split: str | None = None

    results: NestedDict
    meters: AverageMeters
    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  # type: ignore[assignment]
        # 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

    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.
        """

        os.makedirs(os.path.dirname(self.log_path), exist_ok=True)
        # 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_path,
                        "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.

        Notes
        -----
        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:
                    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 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 avoids same data augmentation are applied 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]
        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 update(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, epoch: int | None = None) -> None:
        r"""
        Save checkpoint to `self.checkpoint_dir`.

        The checkpoint will be saved to `self.checkpoint_dir/latest.pth`.

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

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

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

    def load_checkpoint(
        self,
        checkpoint: Mapping | bytes | str | os.PathLike | None = None,
        auto_resume: bool | None = None,
        override_config: bool = False,
        *args,
        **kwargs,
    ) -> None:
        """
        Load info from checkpoint.

        Args:
            checkpoint: Checkpoint (or its path) to load.
                Defaults to `self.config.checkpoint`.
            auto_resume: Automatically resume from latest checkpoint if exists.
                Defaults to `False`.
                If is `True` and `checkpoint` is None, will set it to `self.checkpoint_dir/latest.pth`.
            override_config: If True, override runner config with checkpoint 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.

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

        checkpoint = checkpoint if checkpoint is not None else self.config.get("checkpoint")
        auto_resume = auto_resume if auto_resume is not None else self.config.get("auto_resume", False)

        # TODO: Support loading checkpoints in other format
        if checkpoint is not None:
            if auto_resume:
                warn(
                    "latest checkpoint is preempted by value specified in checkpoint",
                    RuntimeWarning,
                    stacklevel=2,
                )
            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")
                self.config.checkpoint = checkpoint
                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")
        elif auto_resume:
            checkpoint = os.path.join(self.checkpoint_dir, "latest.pth")
            if os.path.exists(checkpoint):
                self.config.checkpoint = checkpoint
                ckpt = self.load(checkpoint, *args, **kwargs)
            else:
                warn("latest checkpoint does not exits", category=RuntimeWarning, stacklevel=2)
                return
        else:
            raise ValueError("checkpoint is not specified and auto_resume is not set to True")

        # TODO: Wrap state_dict in a dataclass
        self.config.merge(ckpt["runner"], overwrite=override_config)
        if self.model is not None and "model" in ckpt:
            model = self.unwrap(self.model)
            model.load_state_dict(ckpt["model"])
        if self.optimizer is not None and "optimizer" in ckpt:
            self.optimizer.load_state_dict(ckpt["optimizer"])
        if self.scheduler is not None and "scheduler" in ckpt:
            self.scheduler.load_state_dict(ckpt["scheduler"])
        self.config.iter_begin = self.config.iter
        self.config.step_begin = self.config.step
        self.config.epoch_begin = self.config.epoch

    @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 | None = None, *args, **kwargs) -> None:
        """
        Load parameters from pretrained checkpoint.

        This method only loads the model weights.

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

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

        See Also:
            [`load_checkpoint`][danling.BaseRunner.load_checkpoint]: Load info from checkpoint.
        """

        # TODO: Support loading checkpoints in other format
        checkpoint = checkpoint if checkpoint is not None else self.config.get("pretrained")
        if checkpoint is None:
            raise ValueError("pretrained is not specified")
        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")
        if self.model is not None and "model" in ckpt:
            model = self.unwrap(self.model)
            model.load_state_dict(ckpt["model"])
        else:
            raise ValueError(f"Unable to find model weights in {checkpoint!r}")

    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.config.epoch
            global __APPEND_RESULT_COUNTER__  # pylint: disable=global-configment
            __APPEND_RESULT_COUNTER__ += 1
            if index == 0 and __APPEND_RESULT_COUNTER__ > 1:
                warn(
                    """
                    Automatically set index to `self.config.epoch`.
                    Please ensure `self.config.epoch` 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.meters.val
        if self.metrics is not None:
            result.merge(self.metrics.val)
        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, step: 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 == "eval":
                repr_str = f"evaluating on {split} "
            else:
                repr_str = f"running in {self.mode} mode on {split} "
        repr_str += f"[{step}/{length}]\t"
        return repr_str + self.format_result(result, format_spec=format_spec)

    def format_epoch_result(
        self, result: NestedDict, epoch: int | None = None, epoch_end: int | None = None, format_spec: str = ".4f"
    ) -> str:
        epoch = epoch or self.config.epoch
        epoch_end = epoch_end or self.config.epoch_end
        repr_str = f"epoch [{epoch}/{epoch_end - 1}]\n" if epoch is not None and epoch_end else ""
        repr_str += "\n".join([f"{k}:\t{self.format_result(v, format_spec=format_spec)}" for k, v in result.items()])
        return repr_str

    def format_result(self, result, format_spec: str = ".4f") -> str:
        result = NestedDict(result).clone()
        return "\t".join([f"{k}: {format(v, format_spec)}" for k, v in result.all_items()])

    def write_result(self, result: NestedDict, split: str, step: int | None = None):
        if step is None:
            step = self.step
        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, step)
            elif isinstance(score, Mapping):
                for k, s in score.items():
                    self.write_score(f"{name}/{k}", s, split, step)
            else:
                self.write_score(name, score, split, step)

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

    @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):
        if "name" in self.config:
            return self.config["name"]
        return f"{self.config.experiment_name}-{self.config.run_name}"

    @cached_property
    def id(self):
        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.

        Notes:
            If `train` is in `dataloaders`, then `batch_size` is the batch size of `train`.
            Otherwise, `batch_size` is the batch size of the first dataloader.

        Returns:
            (int):
        """

        if self.dataloaders and self.split:
            return self.dataloaders[self.split].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"""
        Actual batch size.

        Returns:
            (int): `batch_size` * `world_size` * `accum_steps`
        """

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

    @cached_property
    def total_epochs(self) -> int:
        if self.config.epoch_end:
            return self.config.epoch_end - self.config.epoch_begin + 1
        raise ValueError("epoch_end is not specified")

    @cached_property
    def total_steps(self) -> int:
        if self.config.step_end:
            return self.config.step_end - self.config.step_begin
        dataset = self.datasets.get("train", next(iter(self.datasets.values())))
        return self.total_epochs * ceil(len(dataset) / self.batch_size / self.world_size)

    @cached_property
    def trainable_steps(self) -> int:
        return ceil(self.total_steps / self.accum_steps)

    @cached_property
    def accum_steps(self) -> int:
        r"""
        Accumulated steps.

        Returns:
            (int):
        """

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

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

        Returns:
            (float):

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

        return self.step / self.total_steps

    @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`.

        Returns:
            (callable):
        """

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

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

        Returns:
            (int):
        """

        if not self.scores:
            return 0
        values = list(self.scores.values())
        return self.best_fn(range(len(values)), key=values.__getitem__)

    @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 {IGNORED_SET_NAMES}.
        """

        if not self.results:
            return None
        subsets = [i for i in self.latest_result.keys() if i not in IGNORED_SET_NAMES]  # 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"""
        If current epoch is the best epoch.
        """

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

    @property
    @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_property
    def log_path(self) -> str:
        r"""
        Path of log file.
        """

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

    @property
    @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():
            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

uuid cached property

Python
1
uuid: UUID

UUID of the config.

batch_size property

Python
1
batch_size: int

Batch size.

Notes

If train is in dataloaders, then batch_size is the batch size of train. Otherwise, batch_size is the batch size of the first dataloader.

Returns:

Type Description
int

batch_size_equivalent property

Python
1
batch_size_equivalent: int

Actual batch size.

Returns:

Type Description
int

batch_size * world_size * accum_steps

accum_steps cached property

Python
1
accum_steps: int

Accumulated steps.

Returns:

Type Description
int

progress property

Python
1
progress: float

Training Progress.

Returns:

Type Description
float

Raises:

Type Description
RuntimeError

If no terminal is defined.

world_size property

Python
1
world_size: int

Number of processes.

rank property

Python
1
rank: int

Process index of all processes.

local_rank property

Python
1
local_rank: int

Process index of local processes.

distributed property

Python
1
distributed: bool

If runner is running in distributed mode.

is_main_process property

Python
1
is_main_process: bool

If current process is the main process of all processes.

is_local_main_process property

Python
1
is_local_main_process: bool

If current process is the main process of local processes.

best_fn property

Python
1
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.

Returns:

Type Description
callable

best_index property

Python
1
best_index: int

Find the best index from all scores.

Returns:

Type Description
int

latest_result property

Python
1
latest_result: NestedDict | None

Latest result.

best_result property

Python
1
best_result: NestedDict | None

Best result.

scores property

Python
1
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 {IGNORED_SET_NAMES}.

latest_score property

Python
1
latest_score: float | None

Latest score.

best_score property

Python
1
best_score: float | None

Best score.

is_best property

Python
1
is_best: bool

If current epoch is the best epoch.

dir property

Python
1
dir: str

Directory of the run.

log_path cached property

Python
1
log_path: str

Path of log file.

checkpoint_dir property

Python
1
checkpoint_dir: str

Directory of checkpoints.

init_distributed

Python
1
init_distributed() -> None

Initialise distributed running environment.

Source code in danling/runner/base_runner.py
Python
223
224
225
226
def init_distributed(self) -> None:
    r"""
    Initialise distributed running environment.
    """

init_logging

Python
1
init_logging() -> None

Set up logging.

Source code in danling/runner/base_runner.py
Python
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
@on_main_process
def init_logging(self) -> None:
    r"""
    Set up logging.
    """

    os.makedirs(os.path.dirname(self.log_path), exist_ok=True)
    # 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_path,
                    "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
1
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
Notes

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

Source code in danling/runner/base_runner.py
Python
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
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.

    Notes
    -----
    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:
                logger.info(*args, **kwargs)
            else:
                builtin_print(*args, end=end, file=file, flush=flush, **kwargs)

    __builtin__.print = print

init_tensorboard

Python
1
init_tensorboard(*args, **kwargs) -> None

Set up Tensoraoard SummaryWriter.

Source code in danling/runner/base_runner.py
Python
301
302
303
304
305
306
@on_main_process
def init_tensorboard(self, *args, **kwargs) -> None:
    r"""
    Set up Tensoraoard SummaryWriter.
    """
    raise NotImplementedError

set_seed

Python
1
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 avoids same data augmentation are applied on every processes.

Defaults to self.rank.

Set to False to disable this feature.

 None
Source code in danling/runner/base_runner.py
Python
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
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 avoids same data augmentation are applied 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]
    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
1
set_deterministic() -> None

Set up deterministic.

Source code in danling/runner/base_runner.py
Python
336
337
338
339
340
341
def set_deterministic(self) -> None:
    r"""
    Set up deterministic.
    """

    raise NotImplementedError

scale_lr

Python
1
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
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
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

update

Python
1
update(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
370
371
372
373
374
375
376
377
378
def update(self, loss, *args, **kwargs) -> None:
    r"""
    Backward loss and step optimizer & scheduler.

    Args:
        loss: loss.
    """

    raise NotImplementedError

state_dict

Python
1
state_dict(cls: Callable = dict) -> Mapping

Return dict of all attributes for checkpoint.

Source code in danling/runner/base_runner.py
Python
380
381
382
383
384
385
def state_dict(self, cls: Callable = dict) -> Mapping:
    r"""
    Return dict of all attributes for checkpoint.
    """

    return cls(self.config)

dict

Python
1
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
387
388
389
390
391
392
393
394
395
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
1
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
397
398
399
400
401
402
403
404
405
406
407
408
409
@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
1
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
411
412
413
414
415
416
417
418
419
@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
1
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
421
422
423
424
425
426
427
428
@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
1
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
430
431
432
433
434
435
436
437
438
439
440
@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
1
jsons(*args, **kwargs) -> str

Dump Runner config to json string.

Source code in danling/runner/base_runner.py
Python
442
443
444
445
446
447
def jsons(self, *args, **kwargs) -> str:
    r"""
    Dump Runner config to json string.
    """

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

from_jsons classmethod

Python
1
from_jsons(string: str, *args, **kwargs) -> BaseRunner

Construct Runner from json string.

Source code in danling/runner/base_runner.py
Python
449
450
451
452
453
454
455
@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
1
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
457
458
459
460
461
462
463
464
@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
1
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
466
467
468
469
470
471
472
473
474
475
476
@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
1
yamls(*args, **kwargs) -> str

Dump Runner config to yaml string.

Source code in danling/runner/base_runner.py
Python
478
479
480
481
482
483
def yamls(self, *args, **kwargs) -> str:
    r"""
    Dump Runner config to yaml string.
    """

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

from_yamls classmethod

Python
1
from_yamls(string: str, *args, **kwargs) -> BaseRunner

Construct Runner from yaml string.

Source code in danling/runner/base_runner.py
Python
485
486
487
488
489
490
491
@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
1
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
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
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
1
save_checkpoint(epoch: int | None = None) -> None

Save checkpoint to self.checkpoint_dir.

The checkpoint will be saved to self.checkpoint_dir/latest.pth.

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

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

Source code in danling/runner/base_runner.py
Python
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
@catch
@on_main_process
def save_checkpoint(self, epoch: int | None = None) -> None:
    r"""
    Save checkpoint to `self.checkpoint_dir`.

    The checkpoint will be saved to `self.checkpoint_dir/latest.pth`.

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

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

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

load_checkpoint

Python
1
load_checkpoint(checkpoint: Mapping | bytes | str | PathLike | None = None, auto_resume: bool | None = None, override_config: bool = False, *args, **kwargs) -> None

Load info from checkpoint.

Parameters:

Name Type Description Default
checkpoint Mapping | bytes | str | PathLike | None

Checkpoint (or its path) to load. Defaults to self.config.checkpoint.

 None
auto_resume bool | None

Automatically resume from latest checkpoint if exists. Defaults to False. If is True and checkpoint is None, will set it to self.checkpoint_dir/latest.pth.

 None
override_config bool

If True, override runner config with checkpoint 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.
See Also

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

Source code in danling/runner/base_runner.py
Python
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
def load_checkpoint(
    self,
    checkpoint: Mapping | bytes | str | os.PathLike | None = None,
    auto_resume: bool | None = None,
    override_config: bool = False,
    *args,
    **kwargs,
) -> None:
    """
    Load info from checkpoint.

    Args:
        checkpoint: Checkpoint (or its path) to load.
            Defaults to `self.config.checkpoint`.
        auto_resume: Automatically resume from latest checkpoint if exists.
            Defaults to `False`.
            If is `True` and `checkpoint` is None, will set it to `self.checkpoint_dir/latest.pth`.
        override_config: If True, override runner config with checkpoint 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.

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

    checkpoint = checkpoint if checkpoint is not None else self.config.get("checkpoint")
    auto_resume = auto_resume if auto_resume is not None else self.config.get("auto_resume", False)

    # TODO: Support loading checkpoints in other format
    if checkpoint is not None:
        if auto_resume:
            warn(
                "latest checkpoint is preempted by value specified in checkpoint",
                RuntimeWarning,
                stacklevel=2,
            )
        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")
            self.config.checkpoint = checkpoint
            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")
    elif auto_resume:
        checkpoint = os.path.join(self.checkpoint_dir, "latest.pth")
        if os.path.exists(checkpoint):
            self.config.checkpoint = checkpoint
            ckpt = self.load(checkpoint, *args, **kwargs)
        else:
            warn("latest checkpoint does not exits", category=RuntimeWarning, stacklevel=2)
            return
    else:
        raise ValueError("checkpoint is not specified and auto_resume is not set to True")

    # TODO: Wrap state_dict in a dataclass
    self.config.merge(ckpt["runner"], overwrite=override_config)
    if self.model is not None and "model" in ckpt:
        model = self.unwrap(self.model)
        model.load_state_dict(ckpt["model"])
    if self.optimizer is not None and "optimizer" in ckpt:
        self.optimizer.load_state_dict(ckpt["optimizer"])
    if self.scheduler is not None and "scheduler" in ckpt:
        self.scheduler.load_state_dict(ckpt["scheduler"])
    self.config.iter_begin = self.config.iter
    self.config.step_begin = self.config.step
    self.config.epoch_begin = self.config.epoch

from_checkpoint classmethod

Python
1
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
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
@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
1
load_pretrained(checkpoint: Mapping | bytes | str | PathLike | None = None, *args, **kwargs) -> None

Load parameters from pretrained checkpoint.

This method only loads the model weights.

Parameters:

Name Type Description Default
checkpoint Mapping | bytes | str | PathLike | None

Pretrained checkpoint (or its path) to load. Defaults to self.config.pretrained.

 None
*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.
See Also

load_checkpoint: Load info from checkpoint.

Source code in danling/runner/base_runner.py
Python
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
def load_pretrained(self, checkpoint: Mapping | bytes | str | os.PathLike | None = None, *args, **kwargs) -> None:
    """
    Load parameters from pretrained checkpoint.

    This method only loads the model weights.

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

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

    See Also:
        [`load_checkpoint`][danling.BaseRunner.load_checkpoint]: Load info from checkpoint.
    """

    # TODO: Support loading checkpoints in other format
    checkpoint = checkpoint if checkpoint is not None else self.config.get("pretrained")
    if checkpoint is None:
        raise ValueError("pretrained is not specified")
    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")
    if self.model is not None and "model" in ckpt:
        model = self.unwrap(self.model)
        model.load_state_dict(ckpt["model"])
    else:
        raise ValueError(f"Unable to find model weights in {checkpoint!r}")

append_result

Python
1
append_result(result: NestedDict, index: int | None = None) -> None

Append result to self.results.

Source code in danling/runner/base_runner.py
Python
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
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.config.epoch
        global __APPEND_RESULT_COUNTER__  # pylint: disable=global-configment
        __APPEND_RESULT_COUNTER__ += 1
        if index == 0 and __APPEND_RESULT_COUNTER__ > 1:
            warn(
                """
                Automatically set index to `self.config.epoch`.
                Please ensure `self.config.epoch` 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
1
print_result() -> None

Print latest and best result.

Source code in danling/runner/base_runner.py
Python
704
705
706
707
708
709
710
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
1
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
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
@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)

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
iter int

The number of data samples processed. equals to step when batch_size = 1.
step int

The number of step calls.
epoch int

The number of complete passes over the datasets.
iter_end int

End running iters. Note that step_end not initialised since this variable may not apply to some Runners.
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 only use one of iter_end, step_end, 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.
Notes

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
 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
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:
        iter (int): The number of data samples processed.
            equals to `step` when `batch_size = 1`.
        step (int): The number of `step` calls.
        epoch (int): The number of complete passes over the datasets.
        iter_end (int): End running iters.
            Note that `step_end` not initialised since this variable may not apply to some Runners.
        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 only use one of `iter_end`, `step_end`, `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.

    Notes:
        `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

    iter: int = 0
    step: int = 0
    epoch: int = 0
    iter_begin: int = 0
    step_begin: int = 0
    epoch_begin: int = 0
    iter_end: Optional[int] = None
    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):
        self.copy_class_attributes()
        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_CONFIG_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_CONFIG_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
1
experiment_uuid: UUID

UUID of the experiment.

run_uuid property

Python
1
run_uuid: UUID

UUID of the run.

DeepSpeedRunner

Bases: TorchRunner

Source code in danling/runner/deepspeed_runner.py
Python
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
class DeepSpeedRunner(TorchRunner):

    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.
        """

        torch.cuda.set_device(self.get_local_rank())
        deepspeed.init_distributed()
        object_list = [self.id, self.timestamp]
        dist.broadcast_object_list(object_list)
        self.id, self.timestamp = object_list

    def __post_init__(self):
        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu", self.local_rank)
        super().__post_init__()
        if self.datasets:
            self.build_dataloaders()
        if self.config.get("log_interval") is None:
            self.config.log_interval = max(len(d) for d in self.dataloaders.values()) // 10
        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 backward(self, loss: torch.Tensor) -> None:
        return self.model.backward(loss)

    def advance(self, loss) -> None:
        self.backward(loss)
        self.model.step()

    def get_local_rank(self) -> int:
        local_rank = self.config.get("local_rank", os.getenv("LOCAL_RANK"))
        rank = self.config.get("rank", os.getenv("RANK"))
        world_size = self.config.get("world_size", os.getenv("WORLD_SIZE"))
        if local_rank is None:
            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)
        return int(local_rank)

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

init_distributed

Python
1
init_distributed() -> None

Set up distributed training.

Initialise process group and set up DDP variables.

Source code in danling/runner/deepspeed_runner.py
Python
41
42
43
44
45
46
47
48
49
50
51
52
def init_distributed(self) -> None:
    r"""
    Set up distributed training.

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

    torch.cuda.set_device(self.get_local_rank())
    deepspeed.init_distributed()
    object_list = [self.id, self.timestamp]
    dist.broadcast_object_list(object_list)
    self.id, self.timestamp = object_list

TorchRunner

Bases: BaseRunner

Set up everything for running a job.

TorchRunner uses torch.distributed as distributed backend to provide distributed training experience.

Source code in danling/runner/torch_runner.py
Python
 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
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
class TorchRunner(BaseRunner):
    r"""
    Set up everything for running a job.

    `TorchRunner` uses `torch.distributed` as distributed backend to provide
    distributed training experience.
    """

    model: nn.Module
    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(len(d) for d in self.dataloaders.values()) // 10
        self.model = self.model.to(self.device)
        if self.distributed:
            self.model = nn.parallel.DistributedDataParallel(self.model)

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

        Args:
            train_splits: list of split to run train.
                Defaults to `["train"]`.
            eval_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 eval_splits is None:
            eval_splits = [s for s in self.dataloaders if s not in train_splits]
        self.config.epoch_begin = self.config.epoch
        print(f"Begin training from {self.config.epoch_begin} to {self.config.epoch_end}")
        print(f"Training splits: {train_splits}")
        print(f"Evaluation splits: {eval_splits}")
        patience = self.config.get("patience", float("inf"))
        for epoch in range(self.config.epoch_begin, self.config.epoch_end):  # type: ignore
            self.config.epoch = epoch
            result = NestedDict()
            result.setattr("convert_mapping", True)
            for split in train_splits:
                result[split] = self.train_epoch(split)
            for split in eval_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(epoch)
            """@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.metrics is not None:
            self.metrics.reset()
        batch_time = time()
        if hasattr(loader.batch_sampler, "set_epoch"):
            loader.batch_sampler.set_epoch(self.config.epoch)
        if hasattr(loader.sampler, "set_epoch"):
            loader.sampler.set_epoch(self.config.epoch)

        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.meters.average()
        if self.metrics is not None:
            result.merge(self.metrics.average())
        return result

    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:
            zero_grad: Whether to zero the gradients.

        """

        self.backward(loss / self.accum_steps)
        if self.accum_steps <= 1 or self.config.step % self.accum_steps == 0:
            self.optimizer.step()
            self.optimizer.zero_grad()
        if self.scheduler is not None:
            self.scheduler.step()
        self.config.step += 1

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

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

        Return:
            NestedDict: evaluation result
        """

        if eval_splits is None:
            eval_splits = ["eval"]

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

    @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 = "eval"  # 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.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.meters.average()
        if self.metrics is not None:
            result.merge(self.metrics.average())
        self.write_result(result, split, self.config.epoch)
        return result

    def evaluate_step(self, data) -> torch.Tensor:
        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)
        return loss

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

        Args:
            split (str): split to run inference

        Return:
            Tensor: inference outputs
        """

        self.mode = "inf"  # type: ignore
        loader = self.dataloaders[split]
        self.meters.reset()
        output = []
        for _, data in tqdm(enumerate(loader), total=len(loader)):
            input = data["input"] if isinstance(data, Mapping) else data[0]
            pred = self.model(**input) if isinstance(input, Mapping) else self.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 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 = self.config.get("world_size", os.getenv("WORLD_SIZE"))
        rank = self.config.get("rank", os.getenv("RANK", "0"))
        if world_size is not None and rank is not None:
            dist.init_process_group(backend, init_method, world_size=int(world_size), rank=int(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]
            )

    @staticmethod
    def collate_fn(batch):
        return utils.data.dataloader.default_collate(batch)

    @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.config.step % self.accum_steps == 0:
            yield nullcontext()
        else:
            yield self.model.no_sync()

    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)
        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)

    @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 self.distributed:
            return dist.get_world_size()
        return 1

    @property
    def distributed(self) -> bool:
        return os.getenv("WORLD_SIZE") is not None

    @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
1
world_size: int

Number of Processes.

train

Python
1
train(train_splits: list[str] | None = None, eval_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
eval_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
 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
def train(self, train_splits: list[str] | None = None, eval_splits: list[str] | None = None) -> NestedDict:
    r"""
    Perform training on `split`.

    Args:
        train_splits: list of split to run train.
            Defaults to `["train"]`.
        eval_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 eval_splits is None:
        eval_splits = [s for s in self.dataloaders if s not in train_splits]
    self.config.epoch_begin = self.config.epoch
    print(f"Begin training from {self.config.epoch_begin} to {self.config.epoch_end}")
    print(f"Training splits: {train_splits}")
    print(f"Evaluation splits: {eval_splits}")
    patience = self.config.get("patience", float("inf"))
    for epoch in range(self.config.epoch_begin, self.config.epoch_end):  # type: ignore
        self.config.epoch = epoch
        result = NestedDict()
        result.setattr("convert_mapping", True)
        for split in train_splits:
            result[split] = self.train_epoch(split)
        for split in eval_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(epoch)
        """@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
1
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
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
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.metrics is not None:
        self.metrics.reset()
    batch_time = time()
    if hasattr(loader.batch_sampler, "set_epoch"):
        loader.batch_sampler.set_epoch(self.config.epoch)
    if hasattr(loader.sampler, "set_epoch"):
        loader.sampler.set_epoch(self.config.epoch)

    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.meters.average()
    if self.metrics is not None:
        result.merge(self.metrics.average())
    return result

advance

Python
1
advance(loss) -> None

Backward loss and step optimizer & scheduler.

Parameters:

Name Type Description Default
zero_grad

Whether to zero the gradients.

 required
Source code in danling/runner/torch_runner.py
Python
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
def advance(self, loss) -> None:
    r"""
    Backward loss and step optimizer & scheduler.

    Args:
        zero_grad: Whether to zero the gradients.

    """

    self.backward(loss / self.accum_steps)
    if self.accum_steps <= 1 or self.config.step % self.accum_steps == 0:
        self.optimizer.step()
        self.optimizer.zero_grad()
    if self.scheduler is not None:
        self.scheduler.step()
    self.config.step += 1

evaluate

Python
1
evaluate(eval_splits: list[str] | None = None) -> NestedDict

Perform evaluation on eval_splits.

Parameters:

Name Type Description Default
eval_splits list[str] | None

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

 None
Return
Source code in danling/runner/torch_runner.py
Python
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
def evaluate(self, eval_splits: list[str] | None = None) -> NestedDict:
    r"""
    Perform evaluation on `eval_splits`.

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

    Return:
        NestedDict: evaluation result
    """

    if eval_splits is None:
        eval_splits = ["eval"]

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

evaluate_epoch

Python
1
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
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
@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 = "eval"  # 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.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.meters.average()
    if self.metrics is not None:
        result.merge(self.metrics.average())
    self.write_result(result, split, self.config.epoch)
    return result

inference

Python
1
inference(split: str = 'inf') -> list

Perform inference on split.

Parameters:

Name Type Description Default
split str

split to run inference

 'inf'
Return
Source code in danling/runner/torch_runner.py
Python
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
@torch.inference_mode()
def inference(self, split: str = "inf") -> list:
    r"""
    Perform inference on `split`.

    Args:
        split (str): split to run inference

    Return:
        Tensor: inference outputs
    """

    self.mode = "inf"  # type: ignore
    loader = self.dataloaders[split]
    self.meters.reset()
    output = []
    for _, data in tqdm(enumerate(loader), total=len(loader)):
        input = data["input"] if isinstance(data, Mapping) else data[0]
        pred = self.model(**input) if isinstance(input, Mapping) else self.model(input)
        output.extend(pred.squeeze(-1).tolist())

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

backward

Python
1
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
297
298
299
300
301
302
303
304
305
def backward(self, loss: torch.Tensor) -> None:
    r"""
    Backward loss.

    Args:
        loss: Loss to backward.
    """

    loss.backward()

init_distributed

Python
1
init_distributed() -> None

Set up distributed training.

Initialise process group and set up DDP variables.

Source code in danling/runner/torch_runner.py
Python
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
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 = self.config.get("world_size", os.getenv("WORLD_SIZE"))
    rank = self.config.get("rank", os.getenv("RANK", "0"))
    if world_size is not None and rank is not None:
        dist.init_process_group(backend, init_method, world_size=int(world_size), rank=int(rank))
        object_list = [self.id, self.timestamp]
        dist.broadcast_object_list(object_list)
        self.id, self.timestamp = object_list

init_tensorboard

Python
1
init_tensorboard(*args, **kwargs) -> None

Set up Tensoraoard SummaryWriter.

Source code in danling/runner/torch_runner.py
Python
324
325
326
327
328
329
330
331
332
333
334
335
@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
1
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
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
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
1
get_deepspeed_config(config: NestedDict | str = None) -> NestedDict

Preprocess DeepSpeed config.

Source code in danling/runner/torch_runner.py
Python
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
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)
    return config

NestedTensor

Wrap an iterable of tensors into a single tensor with a mask.

In sequence to sequence tasks, elements of a batch are usually not of the same length. This made it tricky to use a single tensor to represent a batch of sequences.

NestedTensor allows to store a sequence of tensors of different lengths in a single object. It also provides a mask that can be used to retrieve the original sequence of tensors.

When calling __getitem__(arg) on a NestedTensor, it has two return type: 1. if arg is int or slice, returns a tuple of two tensors, representing data and padding mask. 2. if arg is a tuple, return a new NestedTensor with specified shape.

Attributes:

Name Type Description
_storage

The sequence of tensors.
tensor Tensor

padded tensor.
mask Tensor

mask tensor.
concat Tensor

concatenated tensor.
batch_first bool

Whether the first dimension of the tensors is the batch dimension.

If True, the first dimension is the batch dimension, i.e., B, N, *.

If False, the first dimension is the sequence dimension, i.e., N, B, *
padding_value SupportsFloat

The padding value used to in padded tensor.
mask_value bool

The mask value used in mask tensor.

Parameters:

Name Type Description Default
tensors Iterable[Tensor]
()
batch_first bool
True
padding_value SupportsFloat
0.0
mask_value bool
False

Raises:

Type Description
ValueError

If tensors is not an iterable.
ValueError

If tensors is empty.
Notes

We have rewritten the __getattr__ function to support as much native tensor operations as possible. However, not all operations are tested.

Please file an issue if you find any bugs.

Examples:

Python Console Session
>>> nested_tensor = NestedTensor(torch.tensor([1, 2, 3]), torch.tensor([4, 5]))
>>> nested_tensor.shape
torch.Size([2, 3])
>>> nested_tensor.device
device(type='cpu')
>>> nested_tensor.dtype
torch.int64
>>> nested_tensor.tensor
tensor([[1, 2, 3],
        [4, 5, 0]])
>>> nested_tensor.mask
tensor([[ True,  True,  True],
        [ True,  True, False]])
>>> nested_tensor.concat
tensor([1, 2, 3, 4, 5])
>>> 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)
>>> nested_tensor[:]
(tensor([[1, 2, 3],
        [4, 5, 0]]), tensor([[ True,  True,  True],
        [ True,  True, False]]))
>>> nested_tensor[1]
(tensor([4, 5]), tensor([True, True]))
>>> nested_tensor[:, 1:]
NestedTensor([[2, 3],
        [5, 0]])
>>> nested_tensor.tolist()
[[1, 2, 3], [4, 5]]
>>> NestedTensor(*[[1, 2, 3], [4, 5]])
NestedTensor([[1, 2, 3],
        [4, 5, 0]])
Source code in danling/tensors/nested_tensor.py
Python
 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
class NestedTensor:
    r"""
    Wrap an iterable of tensors into a single tensor with a mask.

    In sequence to sequence tasks, elements of a batch are usually not of the same length.
    This made it tricky to use a single tensor to represent a batch of sequences.

    `NestedTensor` allows to store a sequence of tensors of different lengths in a single object.
    It also provides a mask that can be used to retrieve the original sequence of tensors.

    When calling `__getitem__(arg)` on a `NestedTensor`, it has two return type:
    1. if arg is `int` or `slice`, returns a tuple of two `tensor`s, representing data and padding mask.
    2. if arg is a `tuple`, return a new `NestedTensor` with specified shape.

    Attributes:
        _storage: The sequence of tensors.
        tensor: padded tensor.
        mask: mask tensor.
        concat: concatenated tensor.
        batch_first:  Whether the first dimension of the tensors is the batch dimension.

            If `True`, the first dimension is the batch dimension, i.e., `B, N, *`.

            If `False`, the first dimension is the sequence dimension, i.e., `N, B, *`
        padding_value: The padding value used to in padded tensor.
        mask_value: The mask value used in mask tensor.

    Args:
        tensors:
        batch_first:
        padding_value:
        mask_value:

    Raises:
        ValueError: If `tensors` is not an iterable.
        ValueError: If `tensors` is empty.

    Notes:
        We have rewritten the `__getattr__` function to support as much native tensor operations as possible.
        However, not all operations are tested.

        Please file an issue if you find any bugs.

    Examples:
        >>> nested_tensor = NestedTensor(torch.tensor([1, 2, 3]), torch.tensor([4, 5]))
        >>> nested_tensor.shape
        torch.Size([2, 3])
        >>> nested_tensor.device
        device(type='cpu')
        >>> nested_tensor.dtype
        torch.int64
        >>> nested_tensor.tensor
        tensor([[1, 2, 3],
                [4, 5, 0]])
        >>> nested_tensor.mask
        tensor([[ True,  True,  True],
                [ True,  True, False]])
        >>> nested_tensor.concat
        tensor([1, 2, 3, 4, 5])
        >>> 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)
        >>> nested_tensor[:]
        (tensor([[1, 2, 3],
                [4, 5, 0]]), tensor([[ True,  True,  True],
                [ True,  True, False]]))
        >>> nested_tensor[1]
        (tensor([4, 5]), tensor([True, True]))
        >>> nested_tensor[:, 1:]
        NestedTensor([[2, 3],
                [5, 0]])
        >>> nested_tensor.tolist()
        [[1, 2, 3], [4, 5]]
        >>> 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 = [tensor(t) for t in tensors]
        self.__storage = tensors

    def storage(self):
        return self._storage

    @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])
        """
        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.shape[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.view(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 | tuple) -> tuple[Tensor, Tensor] | NestedTensor:
        if isinstance(index, tuple):
            return NestedTensor([t[index[0]][index[1:]] for t in self._storage])
        if isinstance(index, (int, slice)):
            ret = self._storage[index]
            if isinstance(ret, Tensor):
                return ret, torch.ones_like(ret, dtype=torch.bool)
            return self.tensor, self.mask
        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(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)
        size = self.size()
        # ignore channel dimension
        if storage[0].dim() > 1 and len({t.size(-1) for t in storage}) == 1:
            size = size[:-1]  # type: ignore
        return mask_tensor(storage, size=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)

tensor property

Python
1
tensor: Tensor

Return a single tensor by padding all the tensors.

Returns:

Type Description
Tensor

Examples:

Python Console Session
>>> 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
1
mask: Tensor

Padding mask of tensor.

Returns:

Type Description
Tensor

Examples:

Python Console Session
>>> 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
1
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
>>> 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])

device property

Python
1
device: device

Device of the NestedTensor.

Returns:

Type Description
Tensor

Examples:

Python Console Session
>>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
>>> nested_tensor.device
device(type='cpu')

shape property

Python
1
shape: Size | int

Alias for size().

ndim property

Python
1
ndim: int

Alias for dim().

from_tensor_mask classmethod

Python
1
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
>>> 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/tensors/nested_tensor.py
Python
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
@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
1
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
>>> 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/tensors/nested_tensor.py
Python
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
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
1
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
>>> 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/tensors/nested_tensor.py
Python
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
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
1
dim() -> int

Number of dimension of the NestedTensor.

Returns:

Type Description
int

Examples:

Python Console Session
>>> 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/tensors/nested_tensor.py
Python
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
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
1
tolist() -> list

Convert a NestedTensor to a list of lists of values.

Returns:

Type Description
list

Examples:

Python Console Session
>>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
>>> nested_tensor.tolist()
[[1, 2, 3], [4, 5]]
Source code in danling/tensors/nested_tensor.py
Python
452
453
454
455
456
457
458
459
460
461
462
463
464
465
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
1
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
>>> 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/tensors/nested_tensor.py
Python
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
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
1
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
>>> 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/tensors/nested_tensor.py
Python
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
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
1
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
>>> 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/tensors/nested_tensor.py
Python
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
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
1
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
>>> 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/tensors/nested_tensor.py
Python
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
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

Wrapper for tensors to be converted to NestedTensor.

PNTensor is a subclass of torch.Tensor. It implements three additional property as NestedTensor: tensor, mask, and concat.

Although it is possible to directly construct NestedTensor in dataset, the best practice is to do so is in collate_fn. PNTensor is introduced to smoothen the process.

Convert tensors that will be converted to NestedTensor to a PNTensor, and PyTorch Dataloader will automatically collate PNTensor to NestedTensor.

Source code in danling/tensors/nested_tensor.py
Python
 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
class PNTensor(Tensor):
    r"""
    Wrapper for tensors to be converted to `NestedTensor`.

    `PNTensor` is a subclass of `torch.Tensor`.
    It implements three additional property as `NestedTensor`: `tensor`, `mask`, and `concat`.

    Although it is possible to directly construct `NestedTensor` in dataset,
    the best practice is to do so is in `collate_fn`.
    `PNTensor` is introduced to smoothen the process.

    Convert tensors that will be converted to `NestedTensor` to a `PNTensor`,
    and PyTorch Dataloader will automatically collate `PNTensor` to `NestedTensor`.
    """

    @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)

    @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
1
tensor: Tensor

Identical to self.

Returns:

Type Description
Tensor

Examples:

Python Console Session
>>> 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
1
mask: Tensor

Identical to torch.ones_like(self).

Returns:

Type Description
Tensor

Examples:

Python Console Session
>>> 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
1
contact: Tensor

Identical to self.

Returns:

Type Description
Tensor

Examples:

Python Console Session
>>> 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

Metrics

Bases: Metric

Metric class wraps around multiple metrics that share the same states.

Typically, there are many metrics that we want to compute for a single task. For example, we usually needs to compute pearson and spearman for a regression task. Unlike accuracy, which can uses an average meter to compute the average accuracy, pearson and spearman cannot be computed by averaging the results of multiple batches. They need access to all the data to compute the correct results. And saving all intermediate results for each tasks is quite inefficient.

Metrics solves this problem by maintaining a shared state for multiple metric functions.

Attributes:

Name Type Description
metrics FlatDict[str, Callable]

A dictionary of metrics to be computed.
val FlatDict[str, float | flist]

Metric results of current batch on current device.
bat FlatDict[str, float | flist]

Metric results of current batch on all devices.
avg FlatDict[str, float | flist]

Metric results of all results on all devices.
input

The input tensor of latest batch.
target

The target tensor of latest batch.
inputs

All input tensors.
targets

All target tensors.

Parameters:

Name Type Description Default
*args

A single mapping of metrics.

 ()
**metrics Callable

Metrics.

 {}

Examples:

Python Console Session
>>> from danling.metrics.functional import auroc, auprc
>>> 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
FlatDict(
  ('auroc'): 0.75
  ('auprc'): 0.8333333730697632
)
>>> metrics.bat  # Metrics of current batch on all devices
FlatDict(
  ('auroc'): 0.75
  ('auprc'): 0.8333333730697632
)
>>> metrics.avg  # Metrics of all data on all devices
FlatDict(
  ('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
FlatDict(
  ('auroc'): 0.6666666666666666
  ('auprc'): 0.5
)
>>> metrics.bat  # Metrics of current batch on all devices
FlatDict(
  ('auroc'): 0.6666666666666666
  ('auprc'): 0.5
)
>>> metrics.avg  # Metrics of all data on all devices
FlatDict(
  ('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, ignored_index=-100)
>>> 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
(PNTensor([0.1000, 0.6000, 0.8000, 0.1000, 0.6000]), PNTensor([0, 1, 0, 0, 1]))
Source code in danling/metrics/metrics.py
Python
 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
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
class Metrics(Metric):
    r"""
    Metric class wraps around multiple metrics that share the same states.

    Typically, there are many metrics that we want to compute for a single task.
    For example, we usually needs to compute `pearson` and `spearman` for a regression task.
    Unlike `accuracy`, which can uses an average meter to compute the average accuracy,
    `pearson` and `spearman` cannot be computed by averaging the results of multiple batches.
    They need access to all the data to compute the correct results.
    And saving all intermediate results for each tasks is quite inefficient.

    `Metrics` solves this problem by maintaining a shared state for multiple metric functions.

    Attributes:
        metrics: A dictionary of metrics to be computed.
        val: Metric results of current batch on current device.
        bat: Metric results of current batch on all devices.
        avg: Metric results of all results on all devices.
        input: The input tensor of latest batch.
        target: The target tensor of latest batch.
        inputs: All input tensors.
        targets: All target tensors.

    Args:
        *args: A single mapping of metrics.
        **metrics: Metrics.

    Examples:
        >>> from danling.metrics.functional import auroc, auprc
        >>> 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
        FlatDict(
          ('auroc'): 0.75
          ('auprc'): 0.8333333730697632
        )
        >>> metrics.bat  # Metrics of current batch on all devices
        FlatDict(
          ('auroc'): 0.75
          ('auprc'): 0.8333333730697632
        )
        >>> metrics.avg  # Metrics of all data on all devices
        FlatDict(
          ('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
        FlatDict(
          ('auroc'): 0.6666666666666666
          ('auprc'): 0.5
        )
        >>> metrics.bat  # Metrics of current batch on all devices
        FlatDict(
          ('auroc'): 0.6666666666666666
          ('auprc'): 0.5
        )
        >>> metrics.avg  # Metrics of all data on all devices
        FlatDict(
          ('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, ignored_index=-100)
        >>> 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
        (PNTensor([0.1000, 0.6000, 0.8000, 0.1000, 0.6000]), PNTensor([0, 1, 0, 0, 1]))
    """

    metrics: FlatDict[str, Callable]
    ignored_index: Optional[int] = None
    _input: Tensor
    _target: Tensor
    _inputs: flist
    _targets: flist
    _input_buffer: flist
    _target_buffer: flist
    score_name: str
    best_fn: Callable
    merge_dict: bool = True
    return_nested: bool = False

    def __init__(
        self,
        *args,
        merge_dict: bool | None = None,
        return_nested: bool | None = None,
        device: torch.device | None = None,
        ignored_index: int | None = None,
        **metrics: Callable,
    ):
        super().__init__(device=device)
        self._add_state("_input", torch.empty(0))
        self._add_state("_target", torch.empty(0))
        self._add_state("_inputs", flist())
        self._add_state("_targets", flist())
        self._add_state("_input_buffer", flist())
        self._add_state("_target_buffer", flist())
        self.metrics = FlatDict(*args, **metrics)
        if merge_dict is not None:
            self.merge_dict = merge_dict
        if return_nested is not None:
            self.return_nested = return_nested
        self.ignored_index = ignored_index

    @torch.inference_mode()
    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)
        # convert input and target to NestedTensor if one of them is
        if isinstance(input, NestedTensor) or isinstance(target, NestedTensor):
            if isinstance(input, NestedTensor) and isinstance(target, Tensor):
                target = input.nested_like(target, strict=False)
            if isinstance(target, NestedTensor) and isinstance(input, Tensor):
                input = target.nested_like(input, strict=False)
        # remove ignored index
        if self.ignored_index is not None:
            if isinstance(input, NestedTensor):
                indices = [i != self.ignored_index for i in target.storage()]
                input = NestedTensor([t[i] for t, i in zip(input.storage(), indices)])
                target = NestedTensor([t[i] for t, i in zip(target.storage(), indices)])
            else:
                input, target = input[target != self.ignored_index], target[target != self.ignored_index]
        # update internal state
        if isinstance(input, NestedTensor):
            self._input = input
            self._input_buffer.extend(input.cpu().storage())  # type: ignore[union-attr]
            self._target = target
            self._target_buffer.extend(target.cpu().storage())  # type: ignore[union-attr]
        else:
            self._input = input
            self._input_buffer.append(input.cpu())  # type: ignore[union-attr]
            self._target = target
            self._target_buffer.append(target.cpu())  # type: ignore[union-attr]

    def compute(self) -> FlatDict[str, float | flist]:
        return self.calculate(self.inputs.to(self.device), self.targets.to(self.device))

    def value(self) -> FlatDict[str, float | flist]:
        return self.calculate(self._input, self._target)

    def batch(self) -> FlatDict[str, float | flist]:
        return self.calculate(self.input, self.target)

    def average(self) -> FlatDict[str, float | flist]:
        return self.calculate(self.inputs.to(self.device), self.targets.to(self.device))

    @property
    def val(self) -> FlatDict[str, float | flist]:
        return self.value()

    @property
    def bat(self) -> FlatDict[str, float | flist]:
        return self.batch()

    @property
    def avg(self) -> FlatDict[str, float | flist]:
        return self.average()

    @torch.inference_mode()
    def calculate(self, input: Tensor, target: Tensor) -> FlatDict[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 FlatDict({name: nan for name in self.metrics.keys()})
        ret = FlatDict()
        for name, metric in self.metrics.items():
            score = self._calculate(metric, input, target)
            if isinstance(score, Mapping):
                if self.merge_dict:
                    ret.merge(score)
                else:
                    for n, s in score.items():
                        ret[f"{name}.{n}"] = s
            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()

    # Due to an issue with PyTorch, we cannot decorate input/target with @torch.inference_mode()
    # Otherwise, we will encounter the following error when using "gloo" backend:
    # Inplace update to inference tensor outside InferenceMode is not allowed
    @property
    def input(self):
        world_size = get_world_size()
        if world_size == 1:
            if isinstance(self._input, NestedTensor) and not self.return_nested:
                return torch.cat(self._input.storage(), 0)
            return self._input
        if isinstance(self._input, Tensor):
            synced_tensor = [torch.zeros_like(self._input) for _ in range(world_size)]
            dist.all_gather(synced_tensor, self._input)
            return torch.cat(synced_tensor, 0)
        if isinstance(self._input, NestedTensor):
            synced_tensors = [None for _ in range(world_size)]
            dist.all_gather_object(synced_tensors, self._input.storage())
            synced_tensors = flist(i.to(self.device) for j in synced_tensors for i in j)
            try:
                return torch.cat(synced_tensors, 0)
            except RuntimeError:
                if self.return_nested:
                    return NestedTensor(synced_tensors)
                return synced_tensors
        raise ValueError(f"Expected _input to be a Tensor or a NestedTensor, but got {type(self._input)}")

    @property
    def target(self):
        world_size = get_world_size()
        if world_size == 1:
            if isinstance(self._target, NestedTensor) and not self.return_nested:
                return torch.cat(self._target.storage(), 0)
            return self._target
        if isinstance(self._target, Tensor):
            synced_tensor = [torch.zeros_like(self._target) for _ in range(world_size)]
            dist.all_gather(synced_tensor, self._target)
            return torch.cat(synced_tensor, 0)
        if isinstance(self._target, NestedTensor):
            synced_tensors = [None for _ in range(world_size)]
            dist.all_gather_object(synced_tensors, self._target.storage())
            synced_tensors = flist(i.to(self.device) for j in synced_tensors for i in j)
            try:
                return torch.cat(synced_tensors, 0)
            except RuntimeError:
                if self.return_nested:
                    return NestedTensor(synced_tensors)
                return synced_tensors
        raise ValueError(f"Expected _target to be a Tensor or a NestedTensor, but got {type(self._target)}")

    @property
    def inputs(self):
        if not self._inputs and not self._input_buffer:
            return torch.empty(0)
        if self._input_buffer:
            world_size = get_world_size()
            if world_size > 1:
                synced_tensors = [None for _ in range(world_size)]
                dist.all_gather_object(synced_tensors, self._input_buffer)
                self._inputs.extend([i for j in synced_tensors for i in j])
            else:
                self._inputs.extend(self._input_buffer)
            self._input_buffer = flist()
        try:
            return torch.cat(self._inputs, 0)
        except RuntimeError:
            if self.return_nested:
                return NestedTensor(self._inputs)
            return self._inputs

    @property
    def targets(self):
        if not self._targets and not self._target_buffer:
            return torch.empty(0)
        if self._target_buffer:
            world_size = get_world_size()
            if world_size > 1:
                synced_tensors = [None for _ in range(world_size)]
                dist.all_gather_object(synced_tensors, self._target_buffer)
                self._targets.extend([i for j in synced_tensors for i in j])
            else:
                self._targets.extend(self._target_buffer)
            self._target_buffer = flist()
        try:
            return torch.cat(self._targets, 0)
        except RuntimeError:
            if self.return_nested:
                return NestedTensor(self._inputs)
            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

reset

Python
1
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/metrics/metrics.py
Python
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
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

Examples:

Python Console Session
>>> from danling.metrics.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(<class 'danling.metrics.metrics.MultiTaskMetrics'>,
  ('dataset1'): MultiTaskMetrics(<class 'danling.metrics.metrics.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 ...
Source code in danling/metrics/metrics.py
Python
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
class MultiTaskMetrics(MultiTaskDict):
    r"""
    Examples:
        >>> from danling.metrics.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(<class 'danling.metrics.metrics.MultiTaskMetrics'>,
          ('dataset1'): MultiTaskMetrics(<class 'danling.metrics.metrics.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 ...
    """  # 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
1
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/metrics/metrics.py
Python
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
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
1
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
>>> 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
 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
@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
1
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
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
@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

ensure_dir

Python
1
ensure_dir(func)

Decorator to ensure a directory property exists.

Note

Please avoid using this with cached_property.

Examples:

Python Console Session
>>> @property
... @ensure_dir
... def dir(self) -> str:
...     return os.path.join("path", "to", "dir")
Source code in danling/utils/decorators.py
Python
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
def ensure_dir(func):
    r"""
    Decorator to ensure a directory property exists.

    Note:
        Please avoid using this with `cached_property`.

    Examples:
        >>> @property
        ... @ensure_dir
        ... def dir(self) -> str:
        ...     return os.path.join("path", "to", "dir")
    """

    @wraps(func)
    def wrapper(*args, **kwargs):
        path = abspath(func(*args, **kwargs))
        makedirs(path, exist_ok=True)
        return path

    return wrapper

flexible_decorator

Python
1
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
>>> @flexible_decorator
... def decorator(*args, **kwargs):
...     def wrapper(func, *args, **kwargs):
...         pass
...     return wrapper

The following two are equivalent:

Python Console Session
>>> @decorator
... def func(*args, **kwargs):
...     pass
Python Console Session
>>> @decorator()
... def func(*args, **kwargs):
...     pass
Source code in danling/utils/decorators.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
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
1
is_json_serializable(obj: Any) -> bool

Check if obj is JSON serializable.

Source code in danling/utils/io.py
Python
161
162
163
164
165
166
167
168
169
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
1
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
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
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
    if extension in YAML:
        with open(file) as fp:
            return yaml.load(fp, *args, **kwargs)  # type: ignore
    if extension in PICKLE:
        with open(file, "rb") as fp:
            return pickle.load(fp, *args, **kwargs)  # type: ignore
    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
1
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
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
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 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
1
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
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
@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
1
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
 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
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 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
    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
    elif extension in PICKLE:
        with open(file, "wb") as fp:
            pickle.dump(obj, fp, *args, **kwargs)  # type: ignore
    else:
        raise ValueError(f"Tying to save {obj} to {file!r} with unsupported extension={extension!r}")
    return file