Skip to content

Fix PARSeq decoder for TFLite compatibility #2467

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

Merged
sachinprasadhs merged 7 commits into from
Dec 10, 2025
Merged

Fix PARSeq decoder for TFLite compatibility #2467

Changes from 4 commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
3bcafa2
Fix PARSeq decoder for TFLite compatibility
pctablet505 Dec 8, 2025
3a33d0a
Fix PARSeq decoder for TFLite and JAX compatibility
pctablet505 Dec 8, 2025
1cf2750
Fix PARSeq decoder for TFLite and JAX compatibility
pctablet505 Dec 8, 2025
a206ba1
Merge remote-tracking branch 'origin/fix-parseq-decoder-tflite' into …
pctablet505 Dec 8, 2025
eceb781
Update parseq_decoder.py
pctablet505 Dec 8, 2025
620ab07
Update parseq_decoder.py
pctablet505 Dec 9, 2025
e3ae869
Refactor content and query embedding logic in PARSeqDecoder
pctablet505 Dec 10, 2025
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
16 changes: 8 additions & 8 deletions keras_hub/src/models/parseq/parseq_decoder.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -364,14 +364,14 @@ def call(
null_context = self.hidden_dim**0.5 * self.token_embedding(
token_ids[:, :1]
)
if tokens_length > 1:
content = self.pos_query_embeddings[:, : tokens_length - 1, :]
content = content + self.hidden_dim**0.5 * self.token_embedding(
token_ids[:, 1:]
)
content = ops.concatenate([null_context, content], axis=1)
else:
content = null_context

# Build content in a branchless way so both graph and JAX backends
# see a consistent tensor shape. Slicing with length 0 yields empty
# tensors, so concatenation still produces the correct
# (bs, tokens_length, hidden_dim) shape even when tokens_length == 1.
c = self.pos_query_embeddings[:, : tokens_length - 1, :]
c = c + self.hidden_dim**0.5 * self.token_embedding(token_ids[:, 1:])
content = ops.concatenate([null_context, c], axis=1)

content = self.dropout(content)

Expand Down
Loading