phlower.services.trainer.PhlowerTrainer

class phlower.services.trainer.PhlowerTrainer(setting, restart_directory=None)[source]

Bases: object

PhlowerTrainer is a class that manages the training process.

Examples

>>> trainer = PhlowerTrainer.from_setting(setting)
>>> trainer.train(
...     output_directory,
...     train_directories,
...     validation_directories
... )

Methods

__init__(setting[, restart_directory])

Initialize PhlowerTrainer without updating trainer's

attach_handler(name, handler[, allow_overwrite])

Attach handler to the trainer :param name: Name of the handler :type name: str :param handler: Handler to attach :type handler: PhlowerHandlersRunner

draw_model(output_directory)

Draw model

from_setting(setting[, decrypt_key])

Create PhlowerTrainer from PhlowerSetting

get_n_handlers()

Get the number of handlers

get_registered_trainer_setting()

Get registered trainer setting

load_pretrained(model_directory, selection_mode)

Load pretrained model

restart_from(model_directory[, decrypt_key])

Restart PhlowerTrainer from model directory

train()

train_ddp(rank, world_size, output_directory)

Train the model with Distributed Data Parallel (DDP)
Parameters:
attach_handler(name, handler, allow_overwrite=False)[source]

Attach handler to the trainer :param name: Name of the handler :type name: str :param handler: Handler to attach :type handler: PhlowerHandlersRunner

Raises:

ValueError – If handler with the same name already exists

Parameters:

  • name (str)

  • handler (PhlowerHandlersRunner)

  • allow_overwrite (bool)

Return type:

None

draw_model(output_directory)[source]

Draw model

Parameters:

output_directory (Path) – pathlib.Path Output directory

Return type:

None

classmethod from_setting(setting, decrypt_key=None)[source]

Create PhlowerTrainer from PhlowerSetting

Parameters:

  • setting (PhlowerSetting) – PhlowerSetting PhlowerSetting

  • decrypt_key (bytes | None)

Returns:

PhlowerTrainer

Return type:

PhlowerTrainer

get_n_handlers()[source]

Get the number of handlers

Returns:

Number of handlers

Return type:

int

get_registered_trainer_setting()[source]

Get registered trainer setting

Returns:

Trainer setting

Return type:

PhlowerTrainerSetting

load_pretrained(model_directory, selection_mode, target_epoch=None, map_location=None, decrypt_key=None)[source]

Load pretrained model

Parameters:

  • model_directory (Path) – pathlib.Path Model directory

  • selection_mode (Literal['best', 'latest', 'train_best', 'specified']) – Literal[“best”, “latest”, “train_best”, “specified”] Selection mode

  • target_epoch (int | None) – int | None Target epoch. Defaults to None.

  • device – str | None Device. Defaults to None.

  • decrypt_key (bytes | None) – bytes | None Decrypt key. Defaults to None.

  • map_location (str | dict | None)

Return type:

None

classmethod restart_from(model_directory, decrypt_key=None)[source]

Restart PhlowerTrainer from model directory

Parameters:

  • model_directory (Path) – pathlib.Path Model directory

  • decrypt_key (bytes | None)

Returns:

PhlowerTrainer

Return type:

PhlowerTrainer

train_ddp(rank, world_size, output_directory, train_directories=None, validation_directories=None, disable_dimensions=False, decrypt_key=None, encrypt_key=None)[source]

Train the model with Distributed Data Parallel (DDP)

Parameters:

  • rank (int) – Rank of the current process

  • world_size (int) – Total number of processes

  • output_directory (pathlib.Path) – Output directory

  • train_directories (list[pathlib.Path] | None, optional) – List of directories containing training data. If None, directories defined in the setting are used. Default is None.

  • validation_directories (list[pathlib.Path] | None, optional) – List of directories containing validation data. If None, directories defined in the setting are used. Default is None.

  • disable_dimensions (bool, optional) – Disable dimensions. Default is False.

  • decrypt_key (bytes | None, optional) – Key used for decrypting data files, if necessary. Default is None.

  • encrypt_key (bytes | None, optional) – Key used for encrypting output files, if necessary. Default is None.

Return type:

float

Examples

>>> import torch.multiprocessing as mp
>>> trainer = PhlowerTrainer.from_setting(setting)
>>> mp.spawn(
...     trainer.train_ddp,
...     args=(world_size, output_directory),
...     nprocs=world_size,
...     join=True
... )