def __call__(
self,
trainer: GradientDescentTrainer,
metrics: Dict[str, Any],
epoch: int,
is_master: bool,
):
"""Callback call implementation."""
batch = next(iter(trainer._validation_data_loader))
outputs = trainer.model.make_output_human_readable(
trainer.batch_outputs(batch, for_training=False),
)['predicted_sentences']
idx = random.randrange(0, len(outputs))
vocab = trainer.model.vocab
removal_tokens = {START_SYMBOL, END_SYMBOL, vocab._padding_token}
pred_sentence = outputs[idx]
source_sentence = ' '.join(
[
vocab.get_token_from_index(tidx.item())
for tidx in batch['source_tokens']['tokens']['tokens'][idx]
if vocab.get_token_from_index(tidx.item()) not in removal_tokens
],
)
logger.info('{0} -> {1}'.format(source_sentence, pred_sentence))
def __call__(
self,
trainer: GradientDescentTrainer,
batch_inputs: List[List[TensorDict]],
batch_outputs: List[Dict[str, Any]],
epoch: int,
batch_number: int,
is_training: bool,
is_master: bool,
) -> None:
if is_training:
attacker = Attacker(classifier=trainer.model,
reader=self.reader,
device=-1)
for batch in batch_inputs:
instances = []
for element in batch:
data = TransactionsData.from_tensors(
inputs=element, vocab=trainer.model.vocab)
adv_data = attacker.attack(data)
instance = self.reader.text_to_instance(**adv_data)
instances.append(instance)
new_batch = Batch(instances)
new_batch.index_instances(vocab=trainer.model.vocab)
new_batch = new_batch.as_tensor_dict()
batch_outputs = trainer.batch_outputs(new_batch,
for_training=True)
loss = batch_outputs.get("loss")
_ = batch_outputs.get("reg_loss")
loss.backward()
trainer.optimizer.step()
trainer.optimizer.zero_grad()