Skip to content
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

[Fix] PARSeq pytorch fixes #1227

Merged
merged 6 commits into from
Jun 23, 2023
Merged

[Fix] PARSeq pytorch fixes #1227

Changes from 1 commit
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
9ac7c0d
fixes
felixdittrich92 Jun 21, 2023
1b96cc3
up
felixdittrich92 Jun 21, 2023
8e3d42b
update
felixdittrich92 Jun 21, 2023
1cc7fe9
update
felixdittrich92 Jun 21, 2023
f0c1267
update comments and mypy
felixdittrich92 Jun 21, 2023
6755842
update comments and mypy
felixdittrich92 Jun 21, 2023
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
Prev Previous commit
Next Next commit
update
  • Loading branch information
@felixdittrich92
felixdittrich92 committed Jun 21, 2023
commit 8e3d42bc2d6d24f41d528145472a86cba2137fea
16 changes: 4 additions & 12 deletions doctr/models/recognition/parseq/pytorch.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -332,33 +332,25 @@ def forward(
padding_mask = ~(
((gt_in == self.vocab_size + 2) | (gt_in == self.vocab_size)).int().cumsum(-1) > 0
) # (N, seq_len)

key_padding_mask_expanded = padding_mask.view(features.shape[0], 1, 1, gt_in.shape[-1]).expand(
-1, 1, -1, -1
)
) # (N, 1, 1, seq_len)

loss = 0
loss_numel = 0
n = (gt_out != self.vocab_size + 2).sum().item()
for i, perm in enumerate(tgt_perms):
_, target_mask = self.generate_permutations_attention_masks(perm)
# repeat the mask to (batch_size, 1, seq_len, seq_len)
# target_mask = target_mask.unsqueeze(0).repeat(features.shape[0], 1, 1)
target_mask = target_mask.bool()
attn_mask_expanded = target_mask.view(1, 1, gt_in.shape[-1], gt_in.shape[-1]).expand(
features.shape[0], 1, -1, -1
)
) # (N, 1, seq_len, seq_len)
# combine both masks
# mask = (padding_mask.bool() & target_mask.bool()).int()
# tile to (N, 12, seq_len, seq_len)
# mask = mask.unsqueeze(1).repeat(1, 12, 1, 1)
mask = attn_mask_expanded & key_padding_mask_expanded
mask = mask.int()
mask = (attn_mask_expanded.bool() & key_padding_mask_expanded.bool()).int()

logits = self.head(self.decode(gt_out, features, mask)).flatten(end_dim=1)
logits = self.head(self.decode(gt_in, features, mask)).flatten(end_dim=1)
loss += n * F.cross_entropy(logits, gt_out.flatten(), ignore_index=self.vocab_size + 2)
loss_numel += n
# loss += self.compute_loss(logits, gt_out, seq_len, ignore_index=self.vocab_size + 2)
# After the second iteration (i.e. done with canonical and reverse orderings),
# remove the [EOS] tokens for the succeeding perms
if i == 1:
Expand Down