Skip to content

fix: detach probability weights in DFT loss #291

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

Open
taking-lying-flat wants to merge 1 commit into
base: main
Choose a base branch
from
+4 −4
Open

fix: detach probability weights in DFT loss #291

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
6 changes: 3 additions & 3 deletions src/twinkle/loss/chunked_cross_entropy.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -48,7 +48,7 @@ def forward(ctx, logits, labels, chunk_size, ignore_index, reduction, dft):
logps = F.log_softmax(
logits_chunk, dim=-1).gather(-1,
labels_chunk.clamp(min=0).unsqueeze(-1)).squeeze(-1)
per_token = -logps * logps.exp() if dft else -logps
per_token = -logps * logps.exp().detach() if dft else -logps

total_loss = total_loss + (per_token * mask).sum()
total_count = total_count + mask.sum()
Expand Down Expand Up @@ -86,7 +86,7 @@ def backward(ctx, grad_output):
logps = F.log_softmax(
logits_chunk, dim=-1).gather(-1,
labels_chunk.clamp(min=0).unsqueeze(-1)).squeeze(-1)
per_token = -logps * logps.exp() if dft else -logps
per_token = -logps * logps.exp().detach() if dft else -logps
loss_chunk = (per_token * mask).sum()

grad_chunk = torch.autograd.grad(loss_chunk, logits_chunk, retain_graph=False)[0]
Expand Down Expand Up @@ -168,7 +168,7 @@ def __call__(self, inputs, outputs, **kwargs):

def _loss_from_logps(self, labels, logps):
mask = (labels != self.ignore_index).float()
per_token = -logps * logps.exp() if self.dft else -logps
per_token = -logps * logps.exp().detach() if self.dft else -logps
if self.reduction == 'mean':
return LossOutput(loss=(per_token * mask).sum() / mask.sum().clamp(min=1), num_tokens=0)
return LossOutput(loss=(per_token * mask).sum(), num_tokens=mask.sum().clamp(min=1))
2 changes: 1 addition & 1 deletion src/twinkle/loss/cross_entropy.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -39,7 +39,7 @@ def __call__(self, inputs, outputs, **kwargs):

mask = (labels != self.ignore_index).float()
# DFT: -p·log(p) instead of -log(p)
per_token = -logps * logps.exp() if self.dft else -logps
per_token = -logps * logps.exp().detach() if self.dft else -logps

if self.reduction != 'sum':
return LossOutput(loss=(per_token * mask).sum() / mask.sum().clamp(min=1), num_tokens=0)
Expand Down