def concat_all(self,
tensor: Union[Tensor, List[Tensor]],
padding_index: int = -100):
if type(tensor) is list and len(tensor) == 1:
tensor = tensor[0]
self.collector = tensor if self.collector is None else nested_concat(
self.collector, tensor, padding_index)
def prediction_loop(self, data_loader, world_size):
num_examples = len(data_loader.dataset)
batch_size = data_loader.batch_size
eval_losses_gatherer = DistributedTensorGatherer(
world_size, num_examples, make_multiple_of=batch_size)
preds_gatherer = DistributedTensorGatherer(world_size, num_examples)
labels_gatherer = DistributedTensorGatherer(world_size, num_examples)
losses_host, preds_host, labels_host = None, None, None
self.model.eval()
for step, inputs in enumerate(data_loader):
loss, logits, labels = self.prediction_step(inputs)
losses = loss.repeat(batch_size)
losses_host = losses if losses_host is None else torch.cat(
(losses_host, losses), dim=0)
preds_host = logits if preds_host is None else trainer_pt_utils.nested_concat(
preds_host, logits, padding_index=-100)
labels_host = labels if labels_host is None else trainer_pt_utils.nested_concat(
labels_host, labels, padding_index=-100)
eval_losses_gatherer.add_arrays(
trainer_pt_utils.nested_numpify(losses_host))
preds_gatherer.add_arrays(
trainer_pt_utils.nested_numpify(preds_host))
labels_gatherer.add_arrays(
trainer_pt_utils.nested_numpify(labels_host))
losses_host, preds_host, labels_host = None, None, None
eval_loss = eval_losses_gatherer.finalize()
preds = preds_gatherer.finalize()
labels_ids = labels_gatherer.finalize()
if self.type_score == "PER":
preds_ids = np.argmax(preds, axis=-1)
predicted_phonemes = self.processor.batch_decode(
torch.from_numpy(preds_ids))
true_phonemes = self.processor.batch_decode(
torch.from_numpy(labels_ids))
per = generate_per_score(true_phonemes, predicted_phonemes)
return per
elif self.type_score == "WER":
pred = EvalPrediction(predictions=preds, label_ids=labels_ids)
pred_logits = pred.predictions
pred_ids = np.argmax(pred_logits, axis=-1)
pred.label_ids[pred.label_ids ==
-100] = self.processor.tokenizer.pad_token_id
pred_str = self.processor.batch_decode(pred_ids)
# we do not want to group tokens when computing the metrics
label_str = self.processor.batch_decode(pred.label_ids,
group_tokens=False)
metrics = compute_wer(pred_str, label_str)
metrics = denumpify_detensorize(metrics)
metrics["t_loss"] = eval_loss.mean().item()
wer = PredictionOutput(preds, labels_ids, metrics).metrics["wer"]
return wer
start_logits, end_logits = outputs
start_logits = torch.tensor(start_logits)
end_logits = torch.tensor(end_logits)
# necessary to pad predictions and labels for being gathered
start_logits = accelerator.pad_across_processes(start_logits,
dim=1,
pad_index=-100)
end_logits = accelerator.pad_across_processes(end_logits,
dim=1,
pad_index=-100)
logits = (accelerator.gather(start_logits).cpu().numpy(),
accelerator.gather(end_logits).cpu().numpy())
all_preds = logits if all_preds is None else nested_concat(
all_preds, logits, padding_index=-100)
if all_preds is not None:
all_preds = nested_truncate(all_preds, len(eval_dataset))
evalTime = timeit.default_timer() - start_time
logger.info(" Evaluation done in total %f secs (%f sec per example)",
evalTime, evalTime / len(eval_dataset))
# Inference time from TRT
logger.info("Average Inference Time = {:.3f} ms".format(total_time * 1000 /
niter))
logger.info("Total Inference Time = {:.3f} ms".format(total_time * 1000))
logger.info("Total Number of Inference = %d", niter)
prediction = post_processing_function(eval_examples, eval_dataset, all_preds)
eval_metric = metric.compute(predictions=prediction.predictions,
def prediction_loop(
self,
dataloader: DataLoader,
description: str,
prediction_loss_only: Optional[bool] = None,
ignore_keys: Optional[List[str]] = None,
metric_key_prefix: str = "eval",
) -> PredictionOutput:
"""
Prediction/evaluation loop, shared by :obj:`Trainer.evaluate()` and :obj:`Trainer.predict()`.
Works both with or without labels.
"""
if not isinstance(dataloader.dataset, collections.abc.Sized):
raise ValueError("dataset must implement __len__")
prediction_loss_only = (
prediction_loss_only if prediction_loss_only is not None else self.args.prediction_loss_only
)
model = self.model
# multi-gpu eval
if self.args.n_gpu > 1:
model = torch.nn.DataParallel(model)
# Note: in torch.distributed mode, there's no point in wrapping the model
# inside a DistributedDataParallel as we'll be under `no_grad` anyways.
batch_size = dataloader.batch_size
num_examples = self.num_examples(dataloader)
logger.info("***** Running %s *****", description)
logger.info(" Num examples = %d", num_examples)
logger.info(" Batch size = %d", batch_size)
losses_host: torch.Tensor = None
preds_host: Union[torch.Tensor, List[torch.Tensor]] = None
labels_host: Union[torch.Tensor, List[torch.Tensor]] = None
world_size = 1
if is_torch_tpu_available():
world_size = xm.xrt_world_size()
elif self.args.local_rank != -1:
world_size = torch.distributed.get_world_size()
world_size = max(1, world_size)
eval_losses_gatherer = DistributedTensorGatherer(world_size, num_examples, make_multiple_of=batch_size)
if not prediction_loss_only:
preds_gatherer = DistributedTensorGatherer(world_size, num_examples)
labels_gatherer = DistributedTensorGatherer(world_size, num_examples)
model.eval()
if is_torch_tpu_available():
dataloader = pl.ParallelLoader(dataloader, [self.args.device]).per_device_loader(self.args.device)
if self.args.past_index >= 0:
self._past = None
self.callback_handler.eval_dataloader = dataloader
for step, inputs in enumerate(dataloader):
loss, logits, labels = self.prediction_step(model, inputs, prediction_loss_only, ignore_keys=ignore_keys)
if loss is not None:
losses = loss.repeat(batch_size)
losses_host = losses if losses_host is None else torch.cat((losses_host, losses), dim=0)
if logits is not None:
# preds_host = logits if preds_host is None else nested_concat(preds_host, logits, padding_index=-100)
logits_reduced = logits.argmax(-1)
preds_host = logits_reduced if preds_host is None else nested_concat(preds_host, logits_reduced, padding_index=-100)
if labels is not None:
labels_host = labels if labels_host is None else nested_concat(labels_host, labels, padding_index=-100)
self.control = self.callback_handler.on_prediction_step(self.args, self.state, self.control)
# Gather all tensors and put them back on the CPU if we have done enough accumulation steps.
if self.args.eval_accumulation_steps is not None and (step + 1) % self.args.eval_accumulation_steps == 0:
eval_losses_gatherer.add_arrays(self._gather_and_numpify(losses_host, "eval_losses"))
if not prediction_loss_only:
preds_gatherer.add_arrays(self._gather_and_numpify(preds_host, "eval_preds"))
labels_gatherer.add_arrays(self._gather_and_numpify(labels_host, "eval_label_ids"))
# Set back to None to begin a new accumulation
losses_host, preds_host, labels_host = None, None, None
if self.args.past_index and hasattr(self, "_past"):
# Clean the state at the end of the evaluation loop
delattr(self, "_past")
# Gather all remaining tensors and put them back on the CPU
eval_losses_gatherer.add_arrays(self._gather_and_numpify(losses_host, "eval_losses"))
if not prediction_loss_only:
preds_gatherer.add_arrays(self._gather_and_numpify(preds_host, "eval_preds"))
labels_gatherer.add_arrays(self._gather_and_numpify(labels_host, "eval_label_ids"))
eval_loss = eval_losses_gatherer.finalize()
preds = preds_gatherer.finalize() if not prediction_loss_only else None
label_ids = labels_gatherer.finalize() if not prediction_loss_only else None
if self.compute_metrics is not None and preds is not None and label_ids is not None:
metrics = self.compute_metrics(EvalPrediction(predictions=preds, label_ids=label_ids))
else:
metrics = {}
if eval_loss is not None:
metrics[f"{metric_key_prefix}_loss"] = eval_loss.mean().item()
# Prefix all keys with metric_key_prefix + '_'
for key in list(metrics.keys()):
if not key.startswith(f"{metric_key_prefix}_"):
metrics[f"{metric_key_prefix}_{key}"] = metrics.pop(key)
return PredictionOutput(predictions=preds, label_ids=label_ids, metrics=metrics)
def evaluate(self,
dataset,
data_collator=None,
description="",
metric_key_prefix="eval",
compute_metrics=None):
# predicition with single device
eval_sampler = SequentialSampler(dataset)
eval_dataloader = DataLoader(
dataset,
sampler=eval_sampler,
batch_size=self.args.eval_batch_size,
collate_fn=self.data_collator
if data_collator is None else data_collator,
num_workers=self.args.dataloader_num_workers)
batch_size = eval_dataloader.batch_size
num_examples = len(eval_dataloader.dataset)
logger.info("***** Running {} *****".format(description))
logger.info(" Num examples = %d", len(dataset))
logger.info(" Batch size = %d", self.args.eval_batch_size)
losses_host: torch.Tensor = None
preds_host: Union[torch.Tensor, List[torch.Tensor]] = None
labels_host: Union[torch.Tensor, List[torch.Tensor]] = None
world_size = max(1, self.args.world_size)
compute_metrics = self.compute_metrics if compute_metrics is None else compute_metrics
prediction_loss_only = True if compute_metrics is None else None
eval_losses_gatherer = DistributedTensorGatherer(
world_size, num_examples, make_multiple_of=batch_size)
if not prediction_loss_only:
# The actual number of eval_sample can be greater than num_examples in distributed settings (when we pass
# a batch size to the sampler)
make_multiple_of = None
if hasattr(eval_dataloader, "sampler") and isinstance(
eval_dataloader.sampler, SequentialDistributedSampler):
make_multiple_of = eval_dataloader.sampler.batch_size
preds_gatherer = DistributedTensorGatherer(
world_size, num_examples, make_multiple_of=make_multiple_of)
labels_gatherer = DistributedTensorGatherer(
world_size, num_examples, make_multiple_of=make_multiple_of)
model = self._wrap_model(self.model)
model.eval()
all_example_ids = []
start_time = timeit.default_timer()
for step, inputs in enumerate(tqdm(eval_dataloader)):
if 'example_ids' in inputs.keys():
example_ids = inputs.pop('example_ids')
all_example_ids += example_ids
loss, logits, labels = self.prediction_step(
model, inputs, prediction_loss_only)
if loss is not None:
losses = loss.repeat(eval_dataloader.batch_size)
losses_host = losses if losses_host is None else torch.cat(
(losses_host, losses), dim=0)
if logits is not None:
preds_host = logits if preds_host is None else nested_concat(
preds_host, logits, padding_index=-100)
if labels is not None:
labels_host = labels if labels_host is None else nested_concat(
labels_host, labels, padding_index=-100)
# Gather all remaining tensors and put them back on the CPU
eval_losses_gatherer.add_arrays(nested_numpify(losses_host))
if not prediction_loss_only:
preds_gatherer.add_arrays(nested_numpify(preds_host))
labels_gatherer.add_arrays(nested_numpify(labels_host))
eval_loss = eval_losses_gatherer.finalize()
preds = preds_gatherer.finalize() if not prediction_loss_only else None
label_ids = labels_gatherer.finalize(
) if not prediction_loss_only else None
if compute_metrics is not None and preds is not None and label_ids is not None:
metrics = compute_metrics(EvalPrediction(predictions=preds,
label_ids=label_ids),
all_example_ids=all_example_ids
if len(all_example_ids) > 0 else None)
else:
metrics = {}
# To be JSON-serializable, we need to remove numpy types or zero-d tensors
metrics = denumpify_detensorize(metrics)
eval_time = timeit.default_timer() - start_time
logger.info(" Evaluation done in total %f secs (%f sec per example)",
eval_time, eval_time / len(dataset))
if eval_loss is not None:
metrics[f"{metric_key_prefix}_loss"] = eval_loss.mean().item()
# Prefix all keys with metric_key_prefix + '_'
for key in list(metrics.keys()):
if not key.startswith(f"{metric_key_prefix}_"):
metrics[f"{metric_key_prefix}_{key}"] = metrics.pop(key)
return PredictionOutput(
predictions=preds,
label_ids=label_ids,
metrics=metrics,
example_ids=None if len(all_example_ids) == 0 else all_example_ids)
def prediction_loop(
self,
dataloader: DataLoader,
description: str,
prediction_loss_only: Optional[bool] = None,
ignore_keys: Optional[List[str]] = None,
metric_key_prefix: str = "eval",
) -> PredictionOutput:
"""
Prediction/evaluation loop, shared by :obj:`Trainer.evaluate()` and :obj:`Trainer.predict()`.
Works both with or without labels.
"""
if not isinstance(dataloader.dataset, collections.abc.Sized):
raise ValueError("dataset must implement __len__")
prediction_loss_only = (
prediction_loss_only if prediction_loss_only is not None else self.args.prediction_loss_only
)
if self.args.deepspeed and not self.args.do_train:
# no harm, but flagging to the user that deepspeed config is ignored for eval
# flagging only for when --do_train wasn't passed as only then it's redundant
logger.info("Detected the deepspeed argument but it will not be used for evaluation")
model = self.model
# multi-gpu eval
if self.args.n_gpu > 1:
model = torch.nn.DataParallel(model)
# Note: in torch.distributed mode, there's no point in wrapping the model
# inside a DistributedDataParallel as we'll be under `no_grad` anyways
# if full fp16 is wanted on eval and this ``evaluation`` or ``predict`` isn't called while
# ``train`` is running, half it first and then put on device
batch_size = dataloader.batch_size
num_examples = self.num_examples(dataloader)
logger.info("***** Running %s *****", description)
logger.info(" Num examples = %d", num_examples)
logger.info(" Batch size = %d", batch_size)
losses_host: torch.Tensor = None
preds_host: Union[torch.Tensor, List[torch.Tensor]] = None
labels_host: Union[torch.Tensor, List[torch.Tensor]] = None
gumbel_host: Union[torch.Tensor, List[torch.Tensor]] = None
sentence_labels_host: Union[torch.Tensor, List[torch.Tensor]] = None
sentence_indicator_host: Union[torch.Tensor, List[torch.Tensor]] = None
world_size = 1
if is_torch_tpu_available():
world_size = xm.xrt_world_size()
elif self.args.local_rank != -1:
world_size = dist.get_world_size()
world_size = max(1, world_size)
eval_losses_gatherer = DistributedTensorGatherer(world_size, num_examples, make_multiple_of=batch_size)
if not prediction_loss_only:
preds_gatherer = DistributedTensorGatherer(world_size, num_examples)
labels_gatherer = DistributedTensorGatherer(world_size, num_examples)
gumbel_gatherer = DistributedTensorGatherer(world_size, num_examples)
sentence_labels_gatherer = DistributedTensorGatherer(world_size, num_examples)
sentence_indicator_gatherer = DistributedTensorGatherer(world_size, num_examples)
model.eval()
if self.args.past_index >= 0:
self._past = None
self.callback_handler.eval_dataloader = dataloader
for step, inputs in enumerate(dataloader):
loss, logits, labels, gumbel_output, sentence_labels, sentence_indicator = self.prediction_step(model, inputs, prediction_loss_only, ignore_keys=ignore_keys)
if loss is not None:
losses = loss.repeat(batch_size)
losses_host = losses if losses_host is None else torch.cat((losses_host, losses), dim=0)
if logits is not None:
preds_host = logits if preds_host is None else nested_concat(preds_host, logits, padding_index=-100)
if labels is not None:
labels_host = labels if labels_host is None else nested_concat(labels_host, labels, padding_index=-100)
if gumbel_output is not None:
gumbel_host = gumbel_output if gumbel_host is None else nested_concat(gumbel_host, gumbel_output, padding_index=-1)
if sentence_labels is not None:
sentence_labels_host = sentence_labels if sentence_labels_host is None else nested_concat(sentence_labels_host, sentence_labels, padding_index=-1)
if sentence_indicator is not None:
sentence_indicator_host = sentence_indicator if sentence_indicator_host is None else nested_concat(sentence_indicator_host, sentence_indicator, padding_index=-100)
self.control = self.callback_handler.on_prediction_step(self.args, self.state, self.control)
# Gather all tensors and put them back on the CPU if we have done enough accumulation steps.
if self.args.eval_accumulation_steps is not None and (step + 1) % self.args.eval_accumulation_steps == 0:
eval_losses_gatherer.add_arrays(self._gather_and_numpify(losses_host, "eval_losses"))
if not prediction_loss_only:
preds_gatherer.add_arrays(self._gather_and_numpify(preds_host, "eval_preds"))
labels_gatherer.add_arrays(self._gather_and_numpify(labels_host, "eval_label_ids"))
gumbel_gatherer.add_arrays(self._gather_and_numpify(gumbel_host, "eval_gumbel_output"))
sentence_labels_gatherer.add_arrays(self._gather_and_numpify(sentence_labels_host, "eval_sentence_idxs"))
sentence_indicator_gatherer.add_arrays(self._gather_and_numpify(sentence_indicator_host, "eval_sentence_indicator"))
# Set back to None to begin a new accumulation
losses_host, preds_host, labels_host, gumbel_host, sentence_labels_host, sentence_indicator_host = None, None, None, None, None, None
if self.args.past_index and hasattr(self, "_past"):
# Clean the state at the end of the evaluation loop
delattr(self, "_past")
# Gather all remaining tensors and put them back on the CPU
eval_losses_gatherer.add_arrays(self._gather_and_numpify(losses_host, "eval_losses"))
if not prediction_loss_only:
preds_gatherer.add_arrays(self._gather_and_numpify(preds_host, "eval_preds"))
labels_gatherer.add_arrays(self._gather_and_numpify(labels_host, "eval_label_ids"))
gumbel_gatherer.add_arrays(self._gather_and_numpify(gumbel_host, "eval_gumbel_output"))
sentence_labels_gatherer.add_arrays(self._gather_and_numpify(sentence_labels_host, "eval_sentence_idxs"))
sentence_indicator_gatherer.add_arrays(self._gather_and_numpify(sentence_indicator_host, "eval_sentence_indicator"))
eval_loss = eval_losses_gatherer.finalize()
preds = preds_gatherer.finalize() if not prediction_loss_only else None
label_ids = labels_gatherer.finalize() if not prediction_loss_only else None
gumbel_outputs = gumbel_gatherer.finalize() if not prediction_loss_only else None
sentence_idxs = sentence_labels_gatherer.finalize() if not prediction_loss_only else None
sentence_indicators = sentence_indicator_gatherer.finalize() if not prediction_loss_only else None
print(sentence_idxs, 'test')
if self.compute_metrics is not None and preds is not None and label_ids is not None:
metrics = self.compute_metrics(preds, label_ids, gumbel_outputs, sentence_idxs, sentence_indicators)
else:
metrics = {}
if eval_loss is not None:
metrics[f"{metric_key_prefix}_loss"] = eval_loss.mean().item()
# Prefix all keys with metric_key_prefix + '_'
for key in list(metrics.keys()):
if not key.startswith(f"{metric_key_prefix}_"):
metrics[f"{metric_key_prefix}_{key}"] = metrics.pop(key)
return PredictionOutput(predictions=preds, label_ids=label_ids, metrics=metrics)
