V4.22.0 processor update
#2
by
lewtun
HF Staff
- opened
This PR follows #1 and:
- updates the CLIP model to be compatible with
transformersv4.22. The previous version throws an error when trying to load the tokenizer (requiresfrom_slow=True) - sets the image and crop size to the default value of 30 associated with the checkpoint this model was derived from
cc @ydshieh
lewtun
changed pull request title from
V4.22.0 update
to V4.22.0 processor update
The change in tokenizer_config.jsonseems strange to me. It contains crop_size, size, processor_class. I guess it is somehow mixing the fields from preprocessor_config.json?
I generated tokenizer._config.json by using CLIPProcessor.push_to_hub() - presumably these fields are added automatically?
Quick ping on this PR, I encountered this issue while running some TF tests. Is it okay to merge?