2 files changed, 24 insertions, 9 deletions

diff --git a/models/clip/embeddings.py b/models/clip/embeddings.py
index 6be6e9f..8d01867 100644
--- a/models/clip/embeddings.py
+++ b/models/clip/embeddings.py
@@ -38,18 +38,24 @@ class ManagedCLIPTextEmbeddings(CLIPTextEmbeddings):
         self.token_embedding = embeddings.token_embedding
         self.position_embedding = embeddings.position_embedding
         self.initializer_factor = config.initializer_factor
+        self.num_permanent_embeddings = self.token_embedding.num_embeddings
+        self.init_temp_embeddings()
 
+    def init_temp_embeddings(self):
         self.temp_token_embedding = nn.Embedding(
-            self.token_embedding.num_embeddings,
+            0,
             self.token_embedding.embedding_dim,
             device=self.token_embedding.weight.device,
             dtype=self.token_embedding.weight.dtype
         )
-        self.temp_token_embedding.weight.data = self.token_embedding.weight.data.clone().detach()
         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.initializer_factor)
+        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):
@@ -71,7 +77,8 @@ class ManagedCLIPTextEmbeddings(CLIPTextEmbeddings):
         token_ids = torch.tensor(token_ids, dtype=torch.long)
 
         self.temp_token_ids = torch.cat([self.temp_token_ids, token_ids])
-        self.temp_token_embedding.weight.data[token_ids] = initializer.to(
+        mask = torch.nonzero(self.temp_token_ids == token_ids).squeeze(1)
+        self.temp_token_embedding.weight.data[mask] = initializer.to(
             device=self.temp_token_embedding.weight.device,
             dtype=self.temp_token_embedding.weight.dtype,
         )
@@ -85,16 +92,25 @@ class ManagedCLIPTextEmbeddings(CLIPTextEmbeddings):
 
     def persist(self):
         self.token_embedding.weight.data[self.temp_token_ids] = self.temp_token_embedding.weight.data[self.temp_token_ids]
-        self.temp_token_ids = torch.tensor([], dtype=torch.long)
+        self.num_permanent_embeddings = self.token_embedding.num_embeddings
+        self.init_temp_embeddings()
 
     def get_embed(self, input_ids: Union[list[int], torch.LongTensor]):
         if isinstance(input_ids, list):
             input_ids = torch.tensor(input_ids, device=self.token_embedding.weight.device, dtype=torch.long)
 
+        all_temp_token_ids = self.temp_token_ids.to(input_ids.device)
+
         embeds = self.token_embedding(input_ids)
 
-        mask = torch.isin(input_ids, self.temp_token_ids.to(input_ids.device))
-        embeds[mask] = self.temp_token_embedding(input_ids)[mask]
+        embeds_mask = torch.isin(input_ids, all_temp_token_ids)
+        temp_token_ids = input_ids[embeds_mask]
+
+        temp_token_ids = temp_token_ids.unsqueeze(1)
+        all_temp_token_ids = all_temp_token_ids.unsqueeze(0)
+        temp_token_ids = torch.nonzero(temp_token_ids == all_temp_token_ids)[:, 1].squeeze()
+
+        embeds[embeds_mask] = self.temp_token_embedding(temp_token_ids)
 
         return embeds
 
diff --git a/training/strategy/ti.py b/training/strategy/ti.py
index 10bc6d7..b9a5547 100644
--- a/training/strategy/ti.py
+++ b/training/strategy/ti.py
@@ -122,8 +122,7 @@ def textual_inversion_strategy_callbacks(
             if lambda_ != 0:
                 w = text_encoder.text_model.embeddings.temp_token_embedding.weight
 
-                mask = torch.zeros(w.shape[0], dtype=torch.bool)
-                mask[text_encoder.text_model.embeddings.temp_token_ids] = True
+                mask = torch.ones(w.shape[0], dtype=torch.bool)
                 mask[zero_ids] = False
 
                 norm = w[mask, :].norm(dim=-1, keepdim=True)