SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Fix evaluation with label smoothing in Trainer (#10338)

Browse Source
This commit is contained in:
Sylvain Gugger
2021-02-22 16:39:02 -05:00
committed by GitHub
parent 622a8c5995
commit 461e8cacf9
1 changed files with 8 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
src/transformers/trainer.py
View File

@@ -1888,6 +1888,14 @@ class Trainer:
else: else:
ignore_keys = [] ignore_keys = []
# labels may be popped when computing the loss (label smoothing for instance) so we grab them first.
if has_labels:
labels = nested_detach(tuple(inputs.get(name) for name in self.label_names))
if len(labels) == 1:
labels = labels[0]
else:
labels = None
with torch.no_grad(): with torch.no_grad():
if has_labels: if has_labels:
loss, outputs = self.compute_loss(model, inputs, return_outputs=True) loss, outputs = self.compute_loss(model, inputs, return_outputs=True)
@@ -1918,13 +1926,6 @@ class Trainer:
if len(logits) == 1: if len(logits) == 1:
logits = logits[0] logits = logits[0]
if has_labels:
labels = nested_detach(tuple(inputs.get(name) for name in self.label_names))
if len(labels) == 1:
labels = labels[0]
else:
labels = None
return (loss, logits, labels) return (loss, logits, labels)
def floating_point_ops(self, inputs: Dict[str, Union[torch.Tensor, Any]]): def floating_point_ops(self, inputs: Dict[str, Union[torch.Tensor, Any]]):