1 files changed, 16 insertions, 14 deletions

diff --git a/models/clip/embeddings.py b/models/clip/embeddings.py
index 91a575d..cab1515 100644
--- a/models/clip/embeddings.py
+++ b/models/clip/embeddings.py
@@ -12,18 +12,22 @@ from transformers.models.clip import CLIPTextConfig
 from transformers.models.clip.modeling_clip import CLIPTextEmbeddings
 
 
-def expand_embedding(old_embedding: nn.Embedding, n: int) -> nn.Embedding:
+def resize_embedding(old_embedding: nn.Embedding, new_num_embeddings: int, initializer_factor: float = 1.0) -> nn.Embedding:
     old_num_embeddings, old_embedding_dim = old_embedding.weight.size()
 
+    if old_num_embeddings == new_num_embeddings:
+        return old_embedding
+
+    n = min(old_num_embeddings, new_num_embeddings)
+
     new_embedding = nn.Embedding(
-        old_num_embeddings + n,
+        new_num_embeddings,
         old_embedding_dim,
         device=old_embedding.weight.device,
         dtype=old_embedding.weight.dtype
     )
-    new_embedding.weight.data.zero_()
-    new_embedding.weight.data[:old_num_embeddings] = old_embedding.weight.data
-
+    new_embedding.weight.data.normal_(mean=0.0, std=initializer_factor * 0.02)
+    new_embedding.weight.data[:n, :] = old_embedding.weight.data[:n, :]
     return new_embedding
 
 
@@ -40,9 +44,13 @@ class ManagedCLIPTextEmbeddings(CLIPTextEmbeddings):
             device=self.token_embedding.weight.device,
             dtype=self.token_embedding.weight.dtype
         )
-        self.temp_token_embedding.weight.data.zero_()
+        self.temp_token_embedding.weight.data.normal_(mean=0.0, std=config.initializer_factor * 0.02)
         self.temp_token_ids = torch.tensor([], dtype=torch.long)
 
+    def resize(self, size: int):
+        self.temp_token_embedding = resize_embedding(self.temp_token_embedding, size, self.config.initializer_factor)
+        self.token_embedding = resize_embedding(self.token_embedding, size, self.config.initializer_factor)
+
     def add_embed(self, token_ids: Union[int, list[int]], initializer: Optional[Union[int, list[int], torch.FloatTensor]] = None):
         if isinstance(token_ids, int):
             token_ids = [token_ids]
@@ -55,20 +63,14 @@ class ManagedCLIPTextEmbeddings(CLIPTextEmbeddings):
                 initializer = (initializer * len(token_ids))[:len(token_ids)]
 
                 with torch.no_grad():
-                    initializer = self.get_embed(initializer)
-
-        self.temp_token_embedding = expand_embedding(self.temp_token_embedding, len(token_ids))
-        self.token_embedding = expand_embedding(self.token_embedding, len(token_ids))
+                    initializer = self.get_embed(initializer).to(dtype=self.temp_token_embedding.weight.dtype)
 
         token_ids = torch.tensor(token_ids, dtype=torch.long)
 
         self.temp_token_ids = torch.cat([self.temp_token_ids, token_ids])
 
         if initializer is not None:
-            self.temp_token_embedding.weight.data[token_ids] = initializer.to(
-                dtype=self.temp_token_embedding.weight.dtype)
-        else:
-            self.temp_token_embedding.weight.data[token_ids].zero_()
+            self.temp_token_embedding.weight.data[token_ids] = initializer
 
     def load_embed(self, input_ids: list[int], filename: Path):
         with safe_open(filename, framework="pt", device="cpu") as file: