tfm.nlp.tasks.TranslationTask

View source on GitHub

A single-replica view of training procedure.

Inherits From: Task

tfm.nlp.tasks.TranslationTask(
    params: tfm.core.config_definitions.TaskConfig,
    logging_dir=None,
    name=None
)

Tasks provide artifacts for training/evalution procedures, including loading/iterating over Datasets, initializing the model, calculating the loss and customized metrics with reduction.

params the task configuration instance, which can be any of dataclass, ConfigDict, namedtuple, etc. logging_dir a string pointing to where the model, summaries etc. will be saved. You can also write additional stuff in this directory. name the task name.

logging_dir

task_config

Methods

aggregate_logs

View source

aggregate_logs(
    state=None, step_outputs=None
)

Aggregates over logs returned from a validation step.

build_inputs

View source

build_inputs(
    params: tfm.core.config_definitions.DataConfig,
    input_context: Optional[tf.distribute.InputContext] = None
)

Returns a dataset.

build_losses

View source

build_losses(
    labels, model_outputs, aux_losses=None
) -> tf.Tensor

Standard interface to compute losses.

labels optional label tensors. model_outputs a nested structure of output tensors. aux_losses auxiliary loss tensors, i.e. losses in keras.Model.

build_metrics

View source

build_metrics(
    training: bool = True
)

Gets streaming metrics for training/validation.

build_model

View source

build_model() -> tf.keras.Model

Creates model architecture.

create_optimizer

View source

@classmethod
create_optimizer(
    optimizer_config: tfm.optimization.OptimizationConfig,
    runtime_config: Optional[tfm.core.base_task.RuntimeConfig] = None,
    dp_config: Optional[tfm.core.base_task.DifferentialPrivacyConfig] = None
)

Creates an TF optimizer from configurations.

optimizer_config the parameters of the Optimization settings. runtime_config the parameters of the runtime. dp_config the parameter of differential privacy.

inference_step

View source

inference_step(
    inputs, model: tf.keras.Model
)

Performs the forward step.

With distribution strategies, this method runs on devices.

inputs a dictionary of input tensors. model the keras.Model.

initialize

View source

initialize(
    model: tf.keras.Model
)

[Optional] A callback function used as CheckpointManager's init_fn.

This function will be called when no checkpoint is found for the model. If there is a checkpoint, the checkpoint will be loaded and this function will not be called. You can use this callback function to load a pretrained checkpoint, saved under a directory other than the model_dir.

model The keras.Model built or used by this task.

process_compiled_metrics

View source

process_compiled_metrics(
    compiled_metrics, labels, model_outputs
)

Process and update compiled_metrics.

call when using compile/fit API.

compiled_metrics the compiled metrics (model.compiled_metrics). labels a tensor or a nested structure of tensors. model_outputs a tensor or a nested structure of tensors. For example, output of the keras model built by self.build_model.

process_metrics

View source

process_metrics(
    metrics, labels, model_outputs, **kwargs
)

Process and update metrics.

Called when using custom training loop API.

metrics a nested structure of metrics objects. The return of function self.build_metrics. labels a tensor or a nested structure of tensors. model_outputs a tensor or a nested structure of tensors. For example, output of the keras model built by self.build_model. **kwargs other args.

reduce_aggregated_logs

View source

reduce_aggregated_logs(
    aggregated_logs, global_step=None
)

Optional reduce of aggregated logs over validation steps.

This function reduces aggregated logs at the end of validation, and can be used to compute the final metrics. It runs on CPU and in each eval_end() in base trainer (see eval_end() function in official/core/base_trainer.py).

aggregated_logs Aggregated logs over multiple validation steps. global_step An optional variable of global step.

train_step

View source

train_step(
    inputs,
    model: tf.keras.Model,
    optimizer: tf.keras.optimizers.Optimizer,
    metrics=None
)

Does forward and backward.

With distribution strategies, this method runs on devices.

inputs a dictionary of input tensors. model the model, forward pass definition. optimizer the optimizer for this training step. metrics a nested structure of metrics objects.

validation_step

View source

validation_step(
    inputs, model: tf.keras.Model, metrics=None
)

Validation step.

With distribution strategies, this method runs on devices.

inputs a dictionary of input tensors. model the keras.Model. metrics a nested structure of metrics objects.

loss 'loss'