Add SigLIP 2 #36323
qubvel
merged
89 commits into
huggingface:main
from
qubvel:add_siglip2
Docs
412af8c5
Inits
15cd4b0d
Auto classes
dea7c1b4
Add siglip base
a3bc5e2d
Add base tests
b39e4e88
Fix Siglip V1 for fix res version
87d3c512
Add image processor
fda45fbf
Update conversion
2cae39f5
Experimenting with vectorized embeddings
8f7b4baa
Fixup
9d9b515c
Add modular Siglip2Processor
5ff1b57c
Add modular configuration
bcc82825
Rename num patches
69a4a82b
Correct image and text features merging
3fd6e296
Working conversion script
0c85c9ae
Refactoring conversion script
5531808b
Remove unused code in conversion script
8a73f320
Shorten dict a bit
d01ed763
Refactoring conversion
1cf65c80
Done conversion refactoring
fd69948a
Fixup
1e20fb95
Modular siglip2
e07fa7d9
Make model exportable and compilable without graph breaks
496e8351
Remove position_ids from image_processor
d01e68e3
REmove position ids from modeling file
875e627c
Update modular
ecfa936a
Type hint
4cbd2a0d
Fixup
8923186f
Set defaults to processor
869b20cd
Add integration test
bf584bdb
Revert spatial shapes back to tensor
90560e84
Change order
9e280b51
Fix most of the tests
035505bc
Fix docstring
9dc84e22
Remove interpolate_pos_encoding arg (not needed)
7b75b28a
Update docs
424fd333
Standardize processing
0968d80c
Fix attention_mask in vision head
fe0342c4
Siglip v1: remove double transpose in FA2
b136aa06
Update modular file
e506a928
Update FA2 test
bab15ea9
Update expected logits
31ee2008
Fix interpolation for siglip2 image processor
cad7aace
Skip init test
0ac7fe9a
Skip dispatch on flash test
a218959b
Fix modeling tests
758414be
Fixup
30182db2
Add dummy objects
272d922a
Fix some docstrings
22484cbd
Add siglip2 in index.md
b125acc0
Fix consistency
af11bd27
Add docs
f66b9488
Remove size and data format
b79986ac
Add image processor tests
645b384f
Fix
f02aeb06
Add fast image processor
1b94a33b
Fix style
df9f2f65
Fix
206faa86
Docs
50157e90
Set lowercase for tokenizer
5ee7726a
Adjust head size for Siglip v1
d7e07acd
Update siglip2 for consistency with siglip1
aceddebc
Update siglip2 conversion
9c31e405
Update pipeline
8ea956d4
Update checkpoints in tests
32fd65fc
Update checkpoint name
b2d405e3
Fix pooling for image classification model
8c20753e
Fix FA2 test
f7b6e694
Update processor
1cebb122
Fix check repo
c81a4fcb
Update docs
99ae8fbd
Fix typos
902f2843
Fix docstring for fast image processor
a24e4fd4
Add siglip2 to FA2 docs
d905431d
Fix fast ip tests
79c65209
Fix constitency
b38db4c3
Fix tokenizer class for siglip v1
3ceeb3d5
Fix missing header
9cac4e22
Refactor scaling for clip, siglip, siglip2
30b686c7
Remove unused imports
56f75798
Make fast IP default for siglip2
dbc9755f
Update docs
ca9a70f6
Update checkpoints
b04cb9ba
Update modular
5cb7f16c
Update paper link
d79bdd76
Merge branch 'main' into add_siglip2
7892f8a5
Fixup
3bb9430b
LysandreJik
approved these changes
on 2025-02-21
Fix name in toctree
b7746648
Fix test
01c0c7f9
Login to write a write a comment.
Login via GitHub