returnn.engine.base

Provides EngineBase.

class returnn.engine.base.EngineBase(config: Config | None = None)

Base class for a backend engine, such as TFEngine.Engine.

Parameters:

config

init_network_from_config(config: Config | None = None)

Initialize network/model

Parameters:

config

init_train_from_config(config: Config | None = None)

Initialize all engine parts needed for training

Parameters:

config

classmethod config_get_final_epoch(config)

Parameters:

config (returnn.config.Config)

Return type:

int

classmethod get_existing_models(config: Config, *, for_training: bool | None = None)

Parameters:

• config

• for_training – if True, will only return models which are suitable for resuming training. E.g. in case of PyTorch, it means that the optimizer state should be present. By default, will be True if the task is “train”.

Returns:

dict epoch -> model filename (without extension)

Return type:

dict[int,str]

classmethod get_start_epoch_no_existing_model(config: Config) → int

Returns:

start epoch if no model exists

classmethod get_epoch_model(config: Config)

Returns:

(epoch, model_filename). epoch is the epoch of the model filename.

Return type:

(int|None, str|None)

classmethod get_train_start_epoch(config: Config) → int

We will always automatically determine the best start (epoch,batch) tuple based on existing model files. This ensures that the files are present and enforces that there are no old outdated files which should be ignored. Note that epochs start at idx 1 and batches at idx 0.

Parameters:

config

Returns:

epoch

classmethod epoch_model_filename(model_filename: str, epoch: int, *, is_pretrain: bool = False) → str

Parameters:

• model_filename

• epoch

• is_pretrain

get_epoch_model_filename(epoch=None)

Parameters:

epoch (int|None)

Returns:

filename, excluding TF specific postfix

Return type:

str

get_epoch_str()

Returns:

e.g. “epoch 3”, or “pretrain epoch 5”

Return type:

str

is_pretrain_epoch(epoch=None)

Parameters:

epoch (int|None)

Returns:

whether this epoch is covered by the pretrain logic

Return type:

bool

is_first_epoch_after_pretrain()

Returns:

whether the current epoch is the first epoch right after pretraining

Return type:

bool

set_epoch(epoch: int)

Set the current epoch.

forward_with_callback(*, dataset: Dataset, callback: ForwardCallbackIface, dataset_init_epoch: bool = True)

Iterate through the dataset, calling forward_step from user config, collecting outputs in rf.get_run_ctx() via mark_as_output calls, and then calling callback for each entry.

Parameters:

• dataset

• callback – see ForwardCallbackIface

• dataset_init_epoch – whether the engine will call dataset.init_seq_order at the beginning, using the current epoch. If False, it assumes that dataset.init_seq_order was already called.

static delete_model(filename)

Parameters:

filename (str)

Returns:

accumulated file-size in bytes of deleted files

Return type:

int

cleanup_old_models(ask_for_confirmation=False)

Parameters:

ask_for_confirmation (bool) – if True, will ask the user interactively to confirm