how to train tutorial_train.py using IPAttnProcessor.

[kingbackyang]

Hello, IP-Adapter is a great work. When I am reproducing the result and train tutorial_train.py, I find that when you use ./all_models/stable-diffusion-v1-5/ as the args.pretrained_model_name_or_path, the cross_attention_dim in Line 300 will None. This will call the original cross attention (AttnProcessor) but not IPAttnProcessor.

Therefor, I want to know that I want to train the IP-Adapter (IPAttnProcessor), how to set args.pretrained_model_name_or_path? Thank you!

[kingbackyang]

Sorry, I found that you use IPAttnProcessor in unet.attn2.