Method _generate_adversarial_examples

textattack/trainer.py:164–251

Generate adversarial examples using attacker.

(self, epoch)

        self._global_step = 0

    def _generate_adversarial_examples(self, epoch):
        """Generate adversarial examples using attacker."""
        assert (
            self.attack is not None
        ), "`attack` is `None` but attempting to generate adversarial examples."
        base_file_name = f"attack-train-{epoch}"
        log_file_name = os.path.join(self.training_args.output_dir, base_file_name)
        logger.info("Attacking model to generate new adversarial training set...")

        if isinstance(self.training_args.num_train_adv_examples, float):
            num_train_adv_examples = math.ceil(
                len(self.train_dataset) * self.training_args.num_train_adv_examples
            )
        else:
            num_train_adv_examples = self.training_args.num_train_adv_examples

        # Use Different AttackArgs based on num_train_adv_examples value.
        # If num_train_adv_examples >= 0 , num_train_adv_examples is
        # set as number of successful examples.
        # If num_train_adv_examples == -1 , num_examples is set to -1 to
        # generate example for all of training data.
        if num_train_adv_examples >= 0:
            attack_args = AttackArgs(
                num_successful_examples=num_train_adv_examples,
                num_examples_offset=0,
                query_budget=self.training_args.query_budget_train,
                shuffle=True,
                parallel=self.training_args.parallel,
                num_workers_per_device=self.training_args.attack_num_workers_per_device,
                disable_stdout=True,
                silent=True,
                log_to_txt=log_file_name + ".txt",
                log_to_csv=log_file_name + ".csv",
            )
        elif num_train_adv_examples == -1:
            # set num_examples when num_train_adv_examples = -1
            attack_args = AttackArgs(
                num_examples=num_train_adv_examples,
                num_examples_offset=0,
                query_budget=self.training_args.query_budget_train,
                shuffle=True,
                parallel=self.training_args.parallel,
                num_workers_per_device=self.training_args.attack_num_workers_per_device,
                disable_stdout=True,
                silent=True,
                log_to_txt=log_file_name + ".txt",
                log_to_csv=log_file_name + ".csv",
            )
        else:
            assert False, "num_train_adv_examples is negative and not equal to -1."

        attacker = Attacker(self.attack, self.train_dataset, attack_args=attack_args)
        results = attacker.attack_dataset()

        attack_types = collections.Counter(r.__class__.__name__ for r in results)
        total_attacks = (
            attack_types["SuccessfulAttackResult"] + attack_types["FailedAttackResult"]
        )