[Mask2Former] Move normalization for numerical stability
#29542
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
amyeroberts
changed the title
Mask2Former] Move normalization for numerical stability
amyeroberts
commented
Mar 8, 2024
amyeroberts
commented
Mar 8, 2024
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
ydshieh
reviewed
Mar 13, 2024
Comment on lines
376
to
382
| cross_entropy_loss_pos = criterion(inputs, torch.ones_like(inputs)) / height_and_width | ||
| cross_entropy_loss_neg = criterion(inputs, torch.zeros_like(inputs)) / height_and_width | ||
|
|
||
| loss_pos = torch.matmul(cross_entropy_loss_pos, labels.T) | ||
| loss_neg = torch.matmul(cross_entropy_loss_neg, (1 - labels).T) | ||
| loss = loss_pos + loss_neg |
There was a problem hiding this comment.
fine for me. But the name might be a bit misleading (I am not 100% sure): does criterion(inputs, torch.ones_like(inputs)) / height_and_width really representing cross_entropy_loss?
I am not checking the full context here, but personally, I might just do something like
cross_entropy_loss_pos = criterion(inputs, torch.ones_like(inputs))
cross_entropy_loss_neg = criterion(inputs, torch.zeros_like(inputs))
loss_pos = torch.matmul(cross_entropy_loss_pos / height_and_width, labels.T)
loss_neg = torch.matmul(cross_entropy_loss_neg / height_and_width, (1 - labels).T)
to avoid (if any) possible confusion.
There was a problem hiding this comment.
Good point! Used your suggestion
amyeroberts
force-pushed
the
add-normalization-before-matmul
branch
from
51acd85
to
1a3c162
Compare
itazap
pushed a commit
that referenced
this pull request
May 14, 2024
* Move normalization for numerical stability * Apply suggestions from code review Remove useless x=x line * PR comment - normalize later to preserve var name meaning
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What does this PR do?
Moving the normalization before the matmul operation makes the calculation more stable and less likely to overflow.
Same differences introduce in #26086 which was closed after becoming stale.