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

Fix memory regression in Seq2Seq example (#9713)

Browse Source 
* Fix memory regression in Seq2Seq example

* Fix test and properly deal with -100

* Easier condition with device safety

* Patch for MBartTokenzierFast
This commit is contained in:
Sylvain Gugger
2021-01-21 12:05:46 -05:00
committed by GitHub
parent a7dabfb3d1
commit 5f80c15ef5
5 changed files with 43 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
tests/test_trainer_utils.py
View File

@@ -71,7 +71,7 @@ class TrainerUtilsTest(unittest.TestCase):
random_logits = torch.randn(4, 5, num_labels)
random_labels = torch.randint(0, num_labels, (4, 5))
loss = torch.nn.functional.cross_entropy(random_logits.view(-1, num_labels), random_labels.view(-1))
model_output = SequenceClassifierOutput(loss=loss, logits=random_logits)
model_output = SequenceClassifierOutput(logits=random_logits)
label_smoothed_loss = LabelSmoother(0.1)(model_output, random_labels)
log_probs = -torch.nn.functional.log_softmax(random_logits, dim=-1)
expected_loss = (1 - epsilon) * loss + epsilon * log_probs.mean()
@@ -83,7 +83,7 @@ class TrainerUtilsTest(unittest.TestCase):
random_labels[2, 3] = -100
loss = torch.nn.functional.cross_entropy(random_logits.view(-1, num_labels), random_labels.view(-1))
model_output = SequenceClassifierOutput(loss=loss, logits=random_logits)
model_output = SequenceClassifierOutput(logits=random_logits)
label_smoothed_loss = LabelSmoother(0.1)(model_output, random_labels)
log_probs = -torch.nn.functional.log_softmax(random_logits, dim=-1)
# Mask the log probs with the -100 labels