Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Reset loss to zero on logging in Trainer to avoid bfloat16 issues #8561

Merged
merged 5 commits into from
Nov 18, 2020
Merged

Reset loss to zero on logging in Trainer to avoid bfloat16 issues #8561

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
c9d7ccf
make tr_loss regular float
bminixhofer Nov 16, 2020
d3e8b82
Revert "make tr_loss regular float"
bminixhofer Nov 16, 2020
142d34d
reset loss at each logging step
bminixhofer Nov 16, 2020
1f40edb
keep track of total loss with _total_loss_scalar
bminixhofer Nov 16, 2020
ce16f5c
add remaining tr_loss at the end
bminixhofer Nov 17, 2020
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
17 changes: 11 additions & 6 deletions src/transformers/trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -727,8 +727,10 @@ def train(self, model_path: Optional[str] = None, trial: Union["optuna.Trial", D
self.state.is_local_process_zero = self.is_local_process_zero()
self.state.is_world_process_zero = self.is_world_process_zero()

# tr_loss is a tensor to avoid synchronization of TPUs through .item()
tr_loss = torch.tensor(0.0).to(self.args.device)
self._logging_loss_scalar = 0
# _total_loss_scalar is updated everytime .item() has to be called on tr_loss and stores the sum of all losses
self._total_loss_scalar = 0.0
self._globalstep_last_logged = 0
self._total_flos = self.state.total_flos
model.zero_grad()
Expand Down Expand Up @@ -843,23 +845,26 @@ def train(self, model_path: Optional[str] = None, trial: Union["optuna.Trial", D
self.log({"total_flos": self.state.total_flos})

self.control = self.callback_handler.on_train_end(self.args, self.state, self.control)
# add remaining tr_loss
self._total_loss_scalar += tr_loss.item()

return TrainOutput(self.state.global_step, tr_loss.item() / self.state.global_step)
return TrainOutput(self.state.global_step, self._total_loss_scalar / self.state.global_step)

def _maybe_log_save_evaluate(self, tr_loss, model, trial, epoch):
if self.control.should_log:
logs: Dict[str, float] = {}
tr_loss_scalar = tr_loss.item()
logs["loss"] = (tr_loss_scalar - self._logging_loss_scalar) / (
self.state.global_step - self._globalstep_last_logged
)
# reset tr_loss to zero
tr_loss -= tr_loss

logs["loss"] = tr_loss_scalar / (self.state.global_step - self._globalstep_last_logged)
# backward compatibility for pytorch schedulers
logs["learning_rate"] = (
self.lr_scheduler.get_last_lr()[0]
if version.parse(torch.__version__) >= version.parse("1.4")
else self.lr_scheduler.get_lr()[0]
)
self._logging_loss_scalar = tr_loss_scalar
self._total_loss_scalar += tr_loss_scalar
self._globalstep_last_logged = self.state.global_step

self.log(logs)
Expand Down