Method evaluate_step

textattack/trainer.py:551–596

Perform a single evaluation step on a batch of inputs. Args: model (:obj:`torch.nn.Module`): Model to train. tokenizer: Tokenizer used to tokenize input text. batch (:obj:`tuple[list[str], torch.Tensor]`): B

(self, model, tokenizer, batch)

        return loss, preds, _targets

    def evaluate_step(self, model, tokenizer, batch):
        """Perform a single evaluation step on a batch of inputs.

        Args:
            model (:obj:`torch.nn.Module`):
                Model to train.
            tokenizer:
                Tokenizer used to tokenize input text.
            batch (:obj:`tuple[list[str], torch.Tensor]`):
                By default, this will be a tuple of input texts and target tensors.

                .. note::
                    If you override the :meth:`get_eval_dataloader` method, then shape/type of :obj:`batch` will depend on how you created your batch.

        Returns:
            :obj:`tuple[torch.Tensor, torch.Tensor]` where

            - **preds**: :obj:`torch.FloatTensor` of model's prediction for the batch.
            - **targets**: :obj:`torch.Tensor` of model's targets (e.g. labels, target values).
        """
        input_texts, targets = batch
        _targets = targets
        targets = targets.to(textattack.shared.utils.device)

        if isinstance(model, transformers.PreTrainedModel):
            input_ids = tokenizer(
                input_texts,
                padding="max_length",
                return_tensors="pt",
                truncation=True,
            )
            input_ids.to(textattack.shared.utils.device)
            logits = model(**input_ids)[0]
        else:
            input_ids = tokenizer(input_texts)
            if not isinstance(input_ids, torch.Tensor):
                input_ids = torch.tensor(input_ids)
            input_ids = input_ids.to(textattack.shared.utils.device)
            logits = model(input_ids)

        if self.task_type == "regression":
            preds = logits
        else:
            preds = logits.argmax(dim=-1)

        return preds.cpu(), _targets

    def train(self):
        """Train the model on given training dataset."""