modify

misc_utils/train_utils.py

@@ -44,7 +44,7 @@ def get_vae(args):
     if args.get('diffusion'):
         if args.diffusion.params.get('base_path'):
             base_path = args.diffusion.params.base_path
-            vae = AutoencoderKL.from_pretrained(os.path.join(base_path, "vae"))
+            vae = AutoencoderKL.from_pretrained(base_path, subfolder="vae")
         return vae
     return None
 

modules/openclip/modules.py

@@ -100,8 +100,8 @@ class FrozenCLIPEmbedder(AbstractEncoder):
         assert layer in self.LAYERS
         
         if base_path:
-            self.tokenizer = CLIPTokenizer.from_pretrained(os.path.join(base_path, 'tokenizer'))
-            self.transformer = CLIPTextModel.from_pretrained(os.path.join(base_path, 'text_encoder'))
+            self.tokenizer = CLIPTokenizer.from_pretrained(base_path, subfolder='tokenizer')
+            self.transformer = CLIPTextModel.from_pretrained(base_path, subfolder='text_encoder')
         else:
             self.tokenizer = CLIPTokenizer.from_pretrained(version)
             self.transformer = CLIPTextModel.from_pretrained(version)