`tlc.integration.hugging_face.trainer`#

A updated 🤗 transformers Trainer using 3LC capabilities

Module Contents#

Classes#

Class	Description
`TLCTrainer`	A drop-in replacement for the 🤗 transformers Trainer.

API#

class tlc.integration.hugging_face.trainer.TLCTrainer(*args: Any, compute_hf_metrics: Callable[[transformers.trainer_utils.EvalPrediction], dict] | None = None, compute_tlc_metrics: Callable[[torch.Tensor | None, torch.Tensor | None], dict[str, tlc.core.builtins.types.MetricData]] | None = None, compute_tlc_metrics_on_train_begin: bool = False, compute_tlc_metrics_on_train_end: bool = False, tlc_metrics_collection_start: int = 0, tlc_metrics_collection_epoch_frequency: int = -1, **kwargs: Any)#

Bases: transformers.Trainer

A drop-in replacement for the 🤗 transformers Trainer.

Adds per-sample metrics collection on both the train and eval datasets every time .evaluate() is called.

To specify what metrics to collect, pass in a function to the compute_tlc_metrics argument that takes in a batch of data and returns a dictionary of per-sample metrics for the batch.

Parameters:

compute_hf_metrics – The function that will be used to compute metrics at evaluation. Must take a [EvalPrediction] and return a dictionary string to metric values. Also called compute_metrics in HF Trainer.
compute_tlc_metrics – A function that takes in a batch of data and returns a dictionary of metrics.
compute_tlc_metrics_on_train_begin – Whether to collect metrics before training starts.
compute_tlc_metrics_on_train_end – Whether to collect metrics after training ends.
tlc_metrics_collection_start – The iteration or epoch to start collecting metrics on. Can be use with eval_strategy as “epochs” or “steps”. If eval_strategy is “steps”, tlc_metrics_collection_start needs to be a multiple of eval_steps.
tlc_metrics_collection_epoch_frequency – The epoch frequency with which to collect metrics. Must be greater than 0 for metrics to be collected during training. Please use eval_steps for “steps” evaluation strategy.

get_current_epoch() → int#

get_current_global_step() → int#

train(*args: Any, **kwargs: Any) → Any#

Overriding original train method to register start the compute_metrics process when needed.

Args: args (Dict[str, Any]): Arguments to pass to the original train method. kwargs (Dict[str, Any], optional): Additional keyword arguments to pass to the original train method.

Overriding original prediction_step method to compute metrics with 3LC.

Backward compatibility with the original prediction_step method is maintained if compute_tlc_metrics is not set or the specific frequency is not hit.

Args: model (nn.Module): The model to evaluate. inputs (Dict[str, Union[torch.Tensor, Any]]): The inputs and targets of the model.

    The dictionary will be unpacked before being fed to the model. Most models expect the targets under the
    argument `labels`. Check your model's documentation for all accepted arguments.
prediction_loss_only (`bool`):
    Whether or not to return the loss only.
ignore_keys (`List[str]`, *optional*):
    A list of keys in the output of your model (if it is a dictionary) that should be ignored when
    gathering predictions.

Returns: Tuple[Optional[torch.Tensor], Optional[torch.Tensor], Optional[torch.Tensor]]: A tuple with the loss, logits and labels (each being optional).

collect_metrics_with_tlc(ignore_keys: list[str] | None = None, metric_key_prefix: str = 'eval') → dict[str, float]#

Method used to run evaluation on both training and evaluation datasets with tlc metrics collection.

Args: ignore_keys (List[str], optional): A list of keys in the output of your model (if it is a dictionary) that should be ignored when gathering predictions. metric_key_prefix (str, optional, defaults to "eval"): An optional prefix to be used as the metrics key prefix. For example the metrics “bleu” will be named “eval_bleu” if the prefix is “eval” (default)

Returns: A dictionary containing the evaluation loss and the potential metrics computed from the predictions. The dictionary also contains the epoch number which comes from the training state.

run_default_evaluate_based_on_evaluation_strategy(evaluation_strategy: transformers.trainer_utils.IntervalStrategy | str) → bool#

Method to know when we should apply the default evaluation behavior of the Trainer or using 3LC evaluation.

Args: evaluation_strategy (IntervalStrategy or str): The evaluation strategy to check if the current frequency is hit

Returns: A boolean indicating if the default evaluation behavior should be applied or not.

evaluate(eval_dataset: torch.utils.data.Dataset | None = None, ignore_keys: list[str] | None = None, metric_key_prefix: str = 'eval', force_collect_metrics: bool = False) → dict[str, float]#

Overloading method to collect metrics with or without 3LC.

Backward compatibility with the original evaluate method is maintained if compute_tlc_metrics is not set or the specific frequency is not hit.

Args: eval_dataset (Dataset, optional): Pass a dataset if you wish to override self.eval_dataset. If it is a [~datasets.Dataset], columns not accepted by the model.forward() method are automatically removed. It must implement the __len__ method. ignore_keys (List[str], optional): A list of keys in the output of your model (if it is a dictionary) that should be ignored when gathering predictions. metric_key_prefix (str, optional, defaults to "eval"): An optional prefix to be used as the metrics key prefix. For example the metrics “bleu” will be named “eval_bleu” if the prefix is “eval” (default) force_collect_metrics (bool, optional, defaults to False): An optional flag to force the collection of metrics even if the evaluation strategy does not require it. Used specifically to force collecting metrics on train begin or on train end.

Returns: A dictionary containing the evaluation loss and the potential metrics computed from the predictions. The dictionary also contains the epoch number which comes from the training state.

tlc.integration.hugging_face.trainer#

Module Contents#

Classes#

API#

`tlc.integration.hugging_face.trainer`#