def get_eval_dataloader(self,
eval_dataset: Optional[Dataset] = None
) -> DataLoader:
if eval_dataset is None and self.eval_dataset is None:
raise ValueError("Trainer: evaluation requires an eval_dataset.")
eval_dataset = eval_dataset if eval_dataset is not None else self.eval_dataset
if is_tpu_available():
sampler = SequentialDistributedSampler(
eval_dataset,
num_replicas=xm.xrt_world_size(),
rank=xm.get_ordinal())
elif self.args.local_rank != -1:
sampler = SequentialDistributedSampler(eval_dataset)
else:
sampler = SequentialSampler(eval_dataset)
data_loader = DataLoader(
eval_dataset,
sampler=sampler,
batch_size=self.args.eval_batch_size,
collate_fn=self.data_collator.collate_batch,
drop_last=self.args.dataloader_drop_last,
)
return data_loader
def is_world_master(self) -> bool:
"""
This will be True only in one process, even in distributed mode,
even when training on multiple machines.
"""
if is_tpu_available():
return xm.is_master_ordinal(local=False)
else:
return self.args.local_rank == -1 or torch.distributed.get_rank(
) == 0
def __init__(
self,
model: PreTrainedModel,
args: TrainingArguments,
data_collator: Optional[DataCollator] = None,
train_dataset: Optional[Dataset] = None,
eval_dataset: Optional[Dataset] = None,
compute_metrics: Optional[Callable[[EvalPrediction], Dict]] = None,
prediction_loss_only=False,
tb_writer: Optional["SummaryWriter"] = None,
optimizers: Tuple[torch.optim.Optimizer,
torch.optim.lr_scheduler.LambdaLR] = None,
):
"""
Trainer is a simple but feature-complete training and eval loop for PyTorch,
optimized for Transformers.
Args:
prediction_loss_only:
(Optional) in evaluation and prediction, only return the loss
"""
self.model = model.to(args.device)
self.args = args
if data_collator is not None:
self.data_collator = data_collator
else:
self.data_collator = DefaultDataCollator()
self.train_dataset = train_dataset
self.eval_dataset = eval_dataset
self.compute_metrics = compute_metrics
self.prediction_loss_only = prediction_loss_only
self.optimizers = optimizers
if tb_writer is not None:
self.tb_writer = tb_writer
elif is_tensorboard_available() and self.is_world_master():
self.tb_writer = SummaryWriter(log_dir=self.args.logging_dir)
if not is_tensorboard_available():
logger.warning(
"You are instantiating a Trainer but Tensorboard is not installed. You should consider installing it."
)
if is_wandb_available():
self._setup_wandb()
else:
logger.info(
"You are instantiating a Trainer but W&B is not installed. To use wandb logging, "
"run `pip install wandb; wandb login` see https://docs.wandb.com/huggingface."
)
set_seed(self.args.seed)
# Create output directory if needed
if self.is_world_master():
os.makedirs(self.args.output_dir, exist_ok=True)
if is_tpu_available():
# Set an xla_device flag on the model's config.
# We'll find a more elegant and not need to do this in the future.
self.model.config.xla_device = True
def save_model(self, output_dir: Optional[str] = None):
"""
Saving best-practices: if you use default names for the model,
you can reload it using from_pretrained().
Will only save from the world_master process (unless in TPUs).
"""
if is_tpu_available():
self._save_tpu(output_dir)
elif self.is_world_master():
self._save(output_dir)
def get_train_dataloader(self) -> DataLoader:
if self.train_dataset is None:
raise ValueError("Trainer: training requires a train_dataset.")
if is_tpu_available():
train_sampler = get_tpu_sampler(self.train_dataset)
else:
train_sampler = (RandomSampler(self.train_dataset)
if self.args.local_rank == -1 else
DistributedSampler(self.train_dataset))
data_loader = DataLoader(
self.train_dataset,
batch_size=self.args.train_batch_size,
sampler=train_sampler,
collate_fn=self.data_collator.collate_batch,
drop_last=self.args.dataloader_drop_last,
)
return data_loader
def get_test_dataloader(self, test_dataset: Dataset) -> DataLoader:
# We use the same batch_size as for eval.
if is_tpu_available():
sampler = SequentialDistributedSampler(
test_dataset,
num_replicas=xm.xrt_world_size(),
rank=xm.get_ordinal())
elif self.args.local_rank != -1:
sampler = SequentialDistributedSampler(test_dataset)
else:
sampler = SequentialSampler(test_dataset)
data_loader = DataLoader(
test_dataset,
sampler=sampler,
batch_size=self.args.eval_batch_size,
collate_fn=self.data_collator.collate_batch,
)
return data_loader
def _prediction_loop(
self,
dataloader: DataLoader,
description: str,
prediction_loss_only: Optional[bool] = None) -> PredictionOutput:
"""
Prediction/evaluation loop, shared by `evaluate()` and `predict()`.
Works both with or without labels.
"""
prediction_loss_only = prediction_loss_only if prediction_loss_only is not None else self.prediction_loss_only
model = self.model
# multi-gpu eval
if self.args.n_gpu > 1:
model = torch.nn.DataParallel(model)
else:
model = self.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
logger.info("***** Running %s *****", description)
logger.info(" Num examples = %d", self.num_examples(dataloader))
logger.info(" Batch size = %d", batch_size)
eval_losses: List[float] = []
preds: torch.Tensor = None
label_ids: torch.Tensor = None
model.eval()
if is_tpu_available():
dataloader = pl.ParallelLoader(
dataloader,
[self.args.device]).per_device_loader(self.args.device)
for inputs in tqdm(dataloader, desc=description):
has_labels = any(
inputs.get(k) is not None
for k in ["labels", "lm_labels", "masked_lm_labels"])
for k, v in inputs.items():
inputs[k] = v.to(self.args.device)
with torch.no_grad():
outputs = model(**inputs)
if has_labels:
step_eval_loss, logits = outputs[:2]
eval_losses += [step_eval_loss.mean().item()]
else:
logits = outputs[0]
if not prediction_loss_only:
if preds is None:
preds = logits.detach()
else:
preds = torch.cat((preds, logits.detach()), dim=0)
if inputs.get("labels") is not None:
if label_ids is None:
label_ids = inputs["labels"].detach()
else:
label_ids = torch.cat(
(label_ids, inputs["labels"].detach()), dim=0)
if self.args.local_rank != -1:
# In distributed mode, concatenate all results from all nodes:
if preds is not None:
preds = self.distributed_concat(
preds, num_total_examples=self.num_examples(dataloader))
if label_ids is not None:
label_ids = self.distributed_concat(
label_ids,
num_total_examples=self.num_examples(dataloader))
elif is_tpu_available():
# tpu-comment: Get all predictions and labels from all worker shards of eval dataset
if preds is not None:
preds = xm.mesh_reduce("eval_preds", preds, torch.cat)
if label_ids is not None:
label_ids = xm.mesh_reduce("eval_label_ids", label_ids,
torch.cat)
# Finally, turn the aggregated tensors into numpy arrays.
if preds is not None:
preds = preds.cpu().numpy()
if label_ids is not None:
label_ids = label_ids.cpu().numpy()
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 len(eval_losses) > 0:
metrics["eval_loss"] = np.mean(eval_losses)
# Prefix all keys with eval_
for key in list(metrics.keys()):
if not key.startswith("eval_"):
metrics[f"eval_{key}"] = metrics.pop(key)
return PredictionOutput(predictions=preds,
label_ids=label_ids,
metrics=metrics)
def is_local_master(self) -> bool:
if is_tpu_available():
return xm.is_master_ordinal(local=True)
else:
return self.args.local_rank in [-1, 0]
def train(self, model_path: Optional[str] = None):
"""
Main training entry point.
Args:
model_path:
(Optional) Local path to model if model to train has been instantiated from a local path
If present, we will try reloading the optimizer/scheduler states from there.
"""
train_dataloader = self.get_train_dataloader()
if self.args.max_steps > 0:
t_total = self.args.max_steps
num_train_epochs = (self.args.max_steps //
(len(train_dataloader) //
self.args.gradient_accumulation_steps) + 1)
else:
t_total = int(
len(train_dataloader) //
self.args.gradient_accumulation_steps *
self.args.num_train_epochs)
num_train_epochs = self.args.num_train_epochs
optimizer, scheduler = self.get_optimizers(num_training_steps=t_total)
# Check if saved optimizer or scheduler states exist
if (model_path is not None
and os.path.isfile(os.path.join(model_path, "optimizer.pt"))
and os.path.isfile(os.path.join(model_path, "scheduler.pt"))):
# Load in optimizer and scheduler states
optimizer.load_state_dict(
torch.load(os.path.join(model_path, "optimizer.pt"),
map_location=self.args.device))
scheduler.load_state_dict(
torch.load(os.path.join(model_path, "scheduler.pt")))
model = self.model
if self.args.fp16:
if not is_apex_available():
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use fp16 training."
)
model, optimizer = amp.initialize(
model, optimizer, opt_level=self.args.fp16_opt_level)
# multi-gpu training (should be after apex fp16 initialization)
if self.args.n_gpu > 1:
model = torch.nn.DataParallel(model)
# Distributed training (should be after apex fp16 initialization)
if self.args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[self.args.local_rank],
output_device=self.args.local_rank,
find_unused_parameters=True,
)
if self.tb_writer is not None:
self.tb_writer.add_text("args", self.args.to_json_string())
self.tb_writer.add_hparams(self.args.to_sanitized_dict(),
metric_dict={})
# Train!
if is_tpu_available():
total_train_batch_size = self.args.train_batch_size * xm.xrt_world_size(
)
else:
total_train_batch_size = (self.args.train_batch_size *
self.args.gradient_accumulation_steps *
(torch.distributed.get_world_size()
if self.args.local_rank != -1 else 1))
logger.info("***** Running training *****")
logger.info(" Num examples = %d", self.num_examples(train_dataloader))
logger.info(" Num Epochs = %d", num_train_epochs)
logger.info(" Instantaneous batch size per device = %d",
self.args.per_device_train_batch_size)
logger.info(
" Total train batch size (w. parallel, distributed & accumulation) = %d",
total_train_batch_size)
logger.info(" Gradient Accumulation steps = %d",
self.args.gradient_accumulation_steps)
logger.info(" Total optimization steps = %d", t_total)
self.global_step = 0
self.epoch = 0
epochs_trained = 0
steps_trained_in_current_epoch = 0
# Check if continuing training from a checkpoint
if model_path is not None:
# set global_step to global_step of last saved checkpoint from model path
try:
self.global_step = int(model_path.split("-")[-1].split("/")[0])
epochs_trained = self.global_step // (
len(train_dataloader) //
self.args.gradient_accumulation_steps)
steps_trained_in_current_epoch = self.global_step % (
len(train_dataloader) //
self.args.gradient_accumulation_steps)
logger.info(
" Continuing training from checkpoint, will skip to saved global_step"
)
logger.info(" Continuing training from epoch %d",
epochs_trained)
logger.info(" Continuing training from global step %d",
self.global_step)
logger.info(
" Will skip the first %d steps in the first epoch",
steps_trained_in_current_epoch)
except ValueError:
self.global_step = 0
logger.info(" Starting fine-tuning.")
tr_loss = 0.0
logging_loss = 0.0
model.zero_grad()
train_iterator = trange(epochs_trained,
int(num_train_epochs),
desc="Epoch",
disable=not self.is_local_master())
for epoch in train_iterator:
if isinstance(train_dataloader, DataLoader) and isinstance(
train_dataloader.sampler, DistributedSampler):
train_dataloader.sampler.set_epoch(epoch)
if is_tpu_available():
parallel_loader = pl.ParallelLoader(
train_dataloader,
[self.args.device]).per_device_loader(self.args.device)
epoch_iterator = tqdm(parallel_loader,
desc="Iteration",
disable=not self.is_local_master())
else:
epoch_iterator = tqdm(train_dataloader,
desc="Iteration",
disable=not self.is_local_master())
for step, inputs in enumerate(epoch_iterator):
# Skip past any already trained steps if resuming training
if steps_trained_in_current_epoch > 0:
steps_trained_in_current_epoch -= 1
continue
tr_loss += self._training_step(model, inputs, optimizer)
if (step + 1) % self.args.gradient_accumulation_steps == 0 or (
# last step in epoch but step is always smaller than gradient_accumulation_steps
len(epoch_iterator) <=
self.args.gradient_accumulation_steps and
(step + 1) == len(epoch_iterator)):
if self.args.fp16:
torch.nn.utils.clip_grad_norm_(
amp.master_params(optimizer),
self.args.max_grad_norm)
else:
torch.nn.utils.clip_grad_norm_(model.parameters(),
self.args.max_grad_norm)
if is_tpu_available():
xm.optimizer_step(optimizer)
else:
optimizer.step()
scheduler.step()
model.zero_grad()
self.global_step += 1
self.epoch = epoch + (step + 1) / len(epoch_iterator)
if (self.args.logging_steps > 0
and self.global_step % self.args.logging_steps
== 0) or (self.global_step == 1
and self.args.logging_first_step):
logs: Dict[str, float] = {}
logs["loss"] = (tr_loss -
logging_loss) / self.args.logging_steps
# backward compatibility for pytorch schedulers
logs["learning_rate"] = (
scheduler.get_last_lr()[0]
if version.parse(torch.__version__) >=
version.parse("1.4") else scheduler.get_lr()[0])
logging_loss = tr_loss
self._log(logs)
if self.args.evaluate_during_training:
self.evaluate()
if self.args.save_steps > 0 and self.global_step % self.args.save_steps == 0:
# In all cases (even distributed/parallel), self.model is always a reference
# to the model we want to save.
if hasattr(model, "module"):
assert model.module is self.model
else:
assert model is self.model
# Save model checkpoint
output_dir = os.path.join(
self.args.output_dir,
f"{PREFIX_CHECKPOINT_DIR}-{self.global_step}")
self.save_model(output_dir)
if self.is_world_master():
self._rotate_checkpoints()
if is_tpu_available():
xm.rendezvous("saving_optimizer_states")
xm.save(optimizer.state_dict(),
os.path.join(output_dir, "optimizer.pt"))
xm.save(scheduler.state_dict(),
os.path.join(output_dir, "scheduler.pt"))
elif self.is_world_master():
torch.save(
optimizer.state_dict(),
os.path.join(output_dir, "optimizer.pt"))
torch.save(
scheduler.state_dict(),
os.path.join(output_dir, "scheduler.pt"))
if self.args.max_steps > 0 and self.global_step > self.args.max_steps:
epoch_iterator.close()
break
if self.args.max_steps > 0 and self.global_step > self.args.max_steps:
train_iterator.close()
break
if self.args.tpu_metrics_debug:
# tpu-comment: Logging debug metrics for PyTorch/XLA (compile, execute times, ops, etc.)
xm.master_print(met.metrics_report())
if self.tb_writer:
self.tb_writer.close()
logger.info(
"\n\nTraining completed. Do not forget to share your model on huggingface.co/models =)\n\n"
)
return TrainOutput(self.global_step, tr_loss / self.global_step)
from tr_trainer_utils import PREFIX_CHECKPOINT_DIR, EvalPrediction, PredictionOutput, TrainOutput
from tr_training_args import TrainingArguments, is_tpu_available
try:
from apex import amp
_has_apex = True
except ImportError:
_has_apex = False
def is_apex_available():
return _has_apex
if is_tpu_available():
import torch_xla.core.xla_model as xm
import torch_xla.debug.metrics as met
import torch_xla.distributed.parallel_loader as pl
try:
from torch.utils.tensorboard import SummaryWriter
_has_tensorboard = True
except ImportError:
try:
from tensorboardX import SummaryWriter
_has_tensorboard = True
except ImportError:
_has_tensorboard = False