why not use self.clip.transformer when training?

[adda1221]

Hi, thank you for your exciting work! I noticed that you use self.clip.transformer for processing images during the inference stage. However, during the training stage, image processing is accomplished using torchvision. Are there any differences between these two methods? Thanks for your reply!

[Jiaming Han]

Hi @adda1221 , we do not use clip.transformer (the text encoder of clip) during inference. In the training stage, we use simple torchvision transforms which is the same as CLIP's transforms