Skip to content

fix: ignore padding tokens in Bart loss #7828

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

Closed
lvwerra wants to merge 5 commits into from
Closed

fix: ignore padding tokens in Bart loss #7828

Changes from 1 commit
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
0af6246
fix: ignore padding tokens in Bart loss
Oct 15, 2020
21c2205
docs: update Bart model docstring wrt masking the padding token
Oct 15, 2020
a7f3106
test: add assertions for token id -100 in input_ids and labels
Oct 15, 2020
62c6d96
Update src/transformers/modeling_bart.py
sshleifer Oct 15, 2020
9bf323b
Update src/transformers/modeling_bart.py
sshleifer Oct 15, 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
3 changes: 1 addition & 2 deletions src/transformers/modeling_bart.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1110,8 +1110,7 @@ def forward(

masked_lm_loss = None
if labels is not None:
loss_fct = CrossEntropyLoss()
# TODO(SS): do we need to ignore pad tokens in labels?
loss_fct = CrossEntropyLoss(ignore_index=self.config.pad_token_id)

@patil-suraj patil-suraj Oct 15, 2020 •
edited
Loading

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

assert if there are -100 in labels

Wdyt @sshleifer ?

@sshleifer sshleifer Oct 15, 2020

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Great idea. We should also probably do FSMT

masked_lm_loss = loss_fct(lm_logits.view(-1, self.config.vocab_size), labels.view(-1))

if not return_dict:
Expand Down