Method augment

textattack/augmentation/augmenter.py:111–196

Returns all possible augmentations of ``text`` according to ``self.transformation``.

(self, text)

        return transformed_texts

    def augment(self, text):
        """Returns all possible augmentations of ``text`` according to
        ``self.transformation``."""
        attacked_text = AttackedText(text)
        original_text = attacked_text
        all_transformed_texts = set()
        num_words_to_swap = max(
            int(self.pct_words_to_swap * len(attacked_text.words)), 1
        )
        augmentation_results = []
        for _ in range(self.transformations_per_example):
            current_text = attacked_text
            words_swapped = len(current_text.attack_attrs["modified_indices"])

            while words_swapped < num_words_to_swap:
                transformed_texts = self.transformation(
                    current_text, self.pre_transformation_constraints
                )

                # Get rid of transformations we already have
                transformed_texts = [
                    t for t in transformed_texts if t not in all_transformed_texts
                ]

                # Filter out transformations that don't match the constraints.
                transformed_texts = self._filter_transformations(
                    transformed_texts, current_text, original_text
                )

                # if there's no more transformed texts after filter, terminate
                if not len(transformed_texts):
                    break

                # look for all transformed_texts that has enough words swapped
                if self.high_yield or self.fast_augment:
                    ready_texts = [
                        text
                        for text in transformed_texts
                        if len(text.attack_attrs["modified_indices"])
                        >= num_words_to_swap
                    ]
                    for text in ready_texts:
                        all_transformed_texts.add(text)
                    unfinished_texts = [
                        text for text in transformed_texts if text not in ready_texts
                    ]

                    if len(unfinished_texts):
                        current_text = random.choice(unfinished_texts)
                    else:
                        # no need for further augmentations if all of transformed_texts meet `num_words_to_swap`
                        break
                else:
                    current_text = random.choice(transformed_texts)

                # update words_swapped based on modified indices
                words_swapped = max(
                    len(current_text.attack_attrs["modified_indices"]),