PaliGemma#1636

Merged

mattdangerw merged 35 commits intomasterfrom

May 21, 2024

Commits on May 21, 2024

Add VIT Encoder

divyashreepathihalli
authored and
mattdangerw
committed
Add MHAPooling layer + end to end ViT model

fchollet
authored and
mattdangerw
committed
Feature/pg gemma changes

VarunS1997
authored and
mattdangerw
committed
Misc fixes

divyashreepathihalli
authored and
mattdangerw
committed
update vit model and add a test for verifying output shape.

divyashreepathihalli
authored and
mattdangerw
committed
Feature/pg gemma changes

authored and
mattdangerw
committed
Vit model weights conversion

divyashreepathihalli
authored and
mattdangerw
committed
Update imports

divyashreepathihalli
authored and
mattdangerw
committed
Add vit attention

divyashreepathihalli
authored and
mattdangerw
committed
add paligemma functional model

divyashreepathihalli
authored and
mattdangerw
committed
Paligemma full model checkpoints conversion script

divyashreepathihalli
authored and
mattdangerw
committed
Fix ViT build issue

divyashreepathihalli
authored and
mattdangerw
committed
Multi modal Refactor for PaliGemma

VarunS1997
authored and
mattdangerw
committed
Export the public API surface
mattdangerw
committed
update image size arg throughout PaliGemma

divyashreepathihalli
authored and
mattdangerw
committed
Update convert_paligemma_checkpoints.py

divyashreepathihalli
authored and
mattdangerw
committed
Renames for consistency
mattdangerw
committed
Update convert_pali_gemma_checkpoints.py

divyashreepathihalli
authored and
mattdangerw
committed
More consistency improvements for PaliGemma
mattdangerw
committed
Do the same scaling in our backbone we do for generate
mattdangerw
committed
Update conversion and add cli arguments

grasskin
authored and
mattdangerw
committed
Add presets

divyashreepathihalli
authored and
mattdangerw
committed
Update pali_gemma_causal_lm_preprocesor.py to default text sequence length

divyashreepathihalli
authored and
mattdangerw
committed
Tokenizer fix

authored and
mattdangerw
committed
Allow fit calls for pali_gemma
mattdangerw
committed
Allow generate on unbatch input
mattdangerw
committed
Remove the score function from pali gemma causal lm
mattdangerw
committed
Greedy sample by default for pali gemma
mattdangerw
committed
Added docstrings for paligemma decoder and backbone

authored and
mattdangerw
committed
Minor fixes for the vit
mattdangerw
committed
Add a response_mask input

committed
Update pali_gemma_presets.py path

divyashreepathihalli
authored and
mattdangerw
committed
Add a tokenizer docstring for pali gemma
mattdangerw
committed
Update pali_gemma_presets.py

divyashreepathihalli
authored and
mattdangerw
committed
More consistent defaults for PaliGemma

mattdangerw
and
divyashreepathihalli
committed