Ed-ivan

while I ran "pip install pytorch-lightning==1.6.5" (higher version may work well ), it works. May be this will help you.

+1

while I notice that in line 212 of diffusion_transformer.py, log_qt=self.q_pred(log_x_t,t) ,which is called by q_posterior(self,log_x_start,log_x_t,t) . I feel a little confused that why is log_x_t is used but not log_x_start...

In DDPM.py/ class LatentDiffusion (line 810) get_time_with_schedule () , which is called by adapters/coadapters.py/ CoAdapter which inherits the class LatentDiffusion , specificly in line 115 . I think that may...

in evaluate.y line 41: if metric=="psnr_unedit_part": if (1-src_mask).sum()==0 or (1-tgt_mask).sum()==0: return "nan" So , I think the item which is nan is can be ignored. If you have some opinion...

I think that is a good question . And I feel a little confused that why the cropped image is resized to [512,512]

Code : !python /content/HairCLIP/mapper/scripts/train.py \ !--exp_dir=/path/to/experiment \ !--latents_train_path=/path/to/train_faces.pt \ !--latents_test_path=/path/to/test_faces.pt \ !--hairstyle_ref_img_train_path=/path/to/celeba_hq_train \ !--hairstyle_ref_img_test_path=/path/to/celeba_hq_val \ !--color_ref_img_train_path=/path/to/celeba_hq_train \ !--color_ref_img_test_path=/path/to/celeba_hq_val \ !--color_ref_img_in_domain_path=/path/to/generated_hair_of_various colors\ !--hairstyle_manipulation_prob=0.5 \ !--color_manipulation_prob=0.2 \ !--both_manipulation_prob=0.27 \ !--hairstyle_text_manipulation_prob=0.5 \ !--color_text_manipulation_prob=0.5...

While the cls _embeddings is used to replace the text-embeddings , which you can find in the replace_clip_embedding.py