Skip to content

Config

danling.runner.config

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
iters int

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

The number of steps calls.
epochs int

The number of complete passes over the datasets.
iter_end int

End running iter. 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
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:
        iters (int): The number of data samples processed.
            equals to `steps` when `batch_size = 1`.
        steps (int): The number of `steps` calls.
        epochs (int): The number of complete passes over the datasets.
        iter_end (int): End running iter.
            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

    iters: int = 0
    steps: int = 0
    epochs: 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):
        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.