Issue with Fine-Tuning on Colab Pro+ for Japanese Data using Japanese Stable CLIP.

#2
by Kohsaku - opened

Thank you for releasing this wonderful module. I am currently attempting to fine-tune it on Japanese data in a Colab Pro+ environment, but I encounter the following error during training. Upon inspecting the code, I notice that the argument logit_scale is being passed to CLIPOutput, but it seems that this argument does not exist. I am using version 4.36.2 of the transformers library. Any guidance or suggestions to resolve this issue would be greatly appreciated.

~/.cache/huggingface/modules/transformers_modules/stabilityai/japanese-stable-clip-vit-l-16/e08c46e8c46311b5bc8d90928558ae130ca732d7/modeling_japanese_stable_clip.py in forward(self, input_ids, pixel_values, attention_mask, position_ids, return_loss, output_attentions, output_hidden_states, return_dict)
    238             return ((loss,) + output) if loss is not None else output
    239 
--> 240         return CLIPOutput(
    241             loss=loss,
    242             logits_per_image=logits_per_image,

TypeError: CLIPOutput.__init__() got an unexpected keyword argument 'logit_scale'

Sign up or log in to comment