Revert to regular embeddings

1 files changed, 15 insertions, 19 deletions

diff --git a/models/clip/embeddings.py b/models/clip/embeddings.py
index 95904cf..2b315c4 100644
--- a/models/clip/embeddings.py
+++ b/models/clip/embeddings.py
@@ -42,16 +42,20 @@ class ManagedCLIPTextEmbeddings(CLIPTextEmbeddings):
         self.init_temp_embeddings()
 
     def init_temp_embeddings(self):
-        self.temp_token_embedding = nn.ParameterList()
+        self.temp_token_embedding = nn.Embedding(
+            0,
+            self.token_embedding.embedding_dim,
+            device=self.token_embedding.weight.device,
+            dtype=self.token_embedding.weight.dtype
+        )
         self.temp_token_ids = torch.tensor([], dtype=torch.long)
 
     def resize(self, size: int):
-        for _ in range(len(self.temp_token_embedding), size):
-            self.temp_token_embedding.append(torch.zeros(
-                self.token_embedding.embedding_dim,
-                device=self.token_embedding.weight.device,
-                dtype=self.token_embedding.weight.dtype,
-            ))
+        self.temp_token_embedding = resize_embedding(
+            self.temp_token_embedding,
+            size - self.num_permanent_embeddings,
+            self.initializer_factor
+        )
         self.token_embedding = resize_embedding(self.token_embedding, size, self.initializer_factor)
 
     def add_embed(self, token_ids: Union[int, list[int]], initializer: Optional[Union[int, list[int], torch.FloatTensor]] = None):
@@ -78,10 +82,8 @@ class ManagedCLIPTextEmbeddings(CLIPTextEmbeddings):
         token_ids = torch.tensor(token_ids, dtype=torch.long)
 
         self.temp_token_ids = torch.cat([self.temp_token_ids, token_ids])
-        mask = torch.nonzero(torch.isin(self.temp_token_ids, token_ids)).squeeze(1)
-
-        for i, id in enumerate(mask):
-            self.temp_token_embedding[id] = initializer[i]
+        mask = torch.nonzero(self.temp_token_ids == token_ids).squeeze(1)
+        self.temp_token_embedding.weight.data[mask] = initializer
 
     def load_embed(self, input_ids: list[int], filename: Path):
         with safe_open(filename, framework="pt", device="cpu") as file:
@@ -91,8 +93,7 @@ class ManagedCLIPTextEmbeddings(CLIPTextEmbeddings):
         save_file({"embed": self.get_embed(input_ids)}, filename)
 
     def persist(self):
-        for id, emb in zip(self.temp_token_ids, self.temp_token_embedding):
-            self.token_embedding.weight.data[id] = emb
+        self.token_embedding.weight.data[self.temp_token_ids] = self.temp_token_embedding.weight.data[self.temp_token_ids]
         self.num_permanent_embeddings = self.token_embedding.num_embeddings
         self.init_temp_embeddings()
 
@@ -111,12 +112,7 @@ class ManagedCLIPTextEmbeddings(CLIPTextEmbeddings):
         all_temp_token_ids = all_temp_token_ids.unsqueeze(0)
         temp_token_ids = torch.nonzero(temp_token_ids == all_temp_token_ids)[:, 1].squeeze()
 
-        if len(temp_token_ids):
-            embeds_override = torch.stack([
-                self.temp_token_embedding[id]
-                for id in temp_token_ids
-            ])
-            embeds[embeds_mask] = embeds_override
+        embeds[embeds_mask] = self.temp_token_embedding(temp_token_ids)
 
         return embeds