Fix merging embeddings (#226)

Fixed merging embeddings based on the changes made in textual inversion. Tested and working. Inverted their logic to prioritize Stable Diffusion implementation over alternatives, but left the option for alternatives to still be used.

configs/stable-diffusion/v1-finetune.yaml

@@ -52,7 +52,7 @@ model:
         ddconfig:
           double_z: true
           z_channels: 4
-          resolution: 512
+          resolution: 256
           in_channels: 3
           out_ch: 3
           ch: 128
@@ -74,7 +74,7 @@ data:
   target: main.DataModuleFromConfig
   params:
     batch_size: 1
-    num_workers: 16
+    num_workers: 2
     wrap: false
     train:
       target: ldm.data.personalized.PersonalizedBase
@@ -105,4 +105,5 @@ lightning:
 
   trainer:
     benchmark: True
-    max_steps: 6100
+    max_steps: 4000
+    

scripts/merge_embeddings.py

@@ -1,4 +1,4 @@
-from ldm.modules.encoders.modules import BERTTokenizer
+from ldm.modules.encoders.modules import FrozenCLIPEmbedder, BERTEmbedder
 from ldm.modules.embedding_manager import EmbeddingManager
 
 import argparse, os
@@ -6,7 +6,7 @@ from functools import partial
 
 import torch
 
-def get_placeholder_loop(placeholder_string, tokenizer):
+def get_placeholder_loop(placeholder_string, embedder, use_bert):
     
     new_placeholder   = None
     
@@ -16,10 +16,36 @@ def get_placeholder_loop(placeholder_string, tokenizer):
         else:
             new_placeholder = input(f"Placeholder string '{new_placeholder}' maps to more than a single token. Please enter another string: ")
 
-        token = tokenizer(new_placeholder)
+        token = get_bert_token_for_string(embedder.tknz_fn, new_placeholder) if use_bert else get_clip_token_for_string(embedder.tokenizer, new_placeholder)
 
+        if token is not None:
+            return new_placeholder, token
+            
+def get_clip_token_for_string(tokenizer, string):
+    batch_encoding = tokenizer(
+        string,
+        truncation=True,
+        max_length=77,
+        return_length=True,
+        return_overflowing_tokens=False,
+        padding="max_length",
+        return_tensors="pt"
+    )
+
+    tokens = batch_encoding["input_ids"]
+
+    if torch.count_nonzero(tokens - 49407) == 2:
+        return tokens[0, 1]
+    
+    return None
+
+def get_bert_token_for_string(tokenizer, string):
+    token = tokenizer(string)
     if torch.count_nonzero(token) == 3:
-            return new_placeholder, token[0, 1]
+        return token[0, 1]
+
+    return None
+
 
 if __name__ == "__main__":
 
@@ -40,10 +66,20 @@ if __name__ == "__main__":
         help="Output path for the merged manager",
     )
 
+    parser.add_argument(
+        "-sd", "--use_bert",
+        action="store_true",
+        help="Flag to denote that we are not merging stable diffusion embeddings"
+    )
+
     args = parser.parse_args()
 
-    tokenizer = BERTTokenizer(vq_interface=False, max_length=77)
-    EmbeddingManager = partial(EmbeddingManager, tokenizer, ["*"])
+    if args.use_bert:
+        embedder = BERTEmbedder(n_embed=1280, n_layer=32).cuda()
+    else:
+        embedder = FrozenCLIPEmbedder().cuda()
+
+    EmbeddingManager = partial(EmbeddingManager, embedder, ["*"])
 
     string_to_token_dict = {}    
     string_to_param_dict = torch.nn.ParameterDict()
@@ -63,7 +99,7 @@ if __name__ == "__main__":
 
                 placeholder_to_src[placeholder_string] = manager_ckpt
             else:
-                new_placeholder, new_token = get_placeholder_loop(placeholder_string, tokenizer)
+                new_placeholder, new_token = get_placeholder_loop(placeholder_string, embedder, use_bert=args.use_bert)
                 string_to_token_dict[new_placeholder] = new_token
                 string_to_param_dict[new_placeholder] = manager.string_to_param_dict[placeholder_string]
 
@@ -77,7 +113,3 @@ if __name__ == "__main__":
 
     print("Managers merged. Final list of placeholders: ")
     print(placeholder_to_src)
-
-
-
-