Update

6 files changed, 20 insertions, 14 deletions

diff --git a/models/clip/embeddings.py b/models/clip/embeddings.py
index 8c3c6d4..afb7430 100644
--- a/models/clip/embeddings.py
+++ b/models/clip/embeddings.py
@@ -79,8 +79,8 @@ class ManagedCLIPTextEmbeddings(CLIPTextEmbeddings):
     def save_embed(self, input_ids: list[int], filename: Path):
         save_file({"embed": self.get_embed(input_ids)}, filename)
 
-    def persist(self):
-        self.token_embedding.persist()
+    def persist(self, clear=False):
+        self.token_embedding.persist(clear)
 
     def get_embed(self, input_ids: Union[list[int], torch.LongTensor]):
         if isinstance(input_ids, list):
diff --git a/models/sparse.py b/models/sparse.py
index e5897c9..55c9837 100644
--- a/models/sparse.py
+++ b/models/sparse.py
@@ -89,10 +89,15 @@ class SparseEmbedding(nn.Embedding):
 
         return weights
 
-    def persist(self):
+    def persist(self, clear=False):
         self.weight.data += self.get_weights(torch.arange(self.trainable_ids.shape[0]))
-        self.trainable_ids[:] = -1
-        self.trainable = nn.ParameterList()
+
+        if clear:
+            self.trainable_ids[:] = -1
+            self.trainable = nn.ParameterList()
+        else:
+            for param in self.trainable:
+                param.zero_()
 
     def reset_parameters(self):
         nn.Embedding.reset_parameters(self)
diff --git a/train_dreambooth.py b/train_dreambooth.py
index beb65fc..929310b 100644
--- a/train_dreambooth.py
+++ b/train_dreambooth.py
@@ -661,7 +661,7 @@ def main():
             placeholder_tokens=alias_placeholder_tokens,
             initializer_tokens=alias_initializer_tokens,
         )
-        embeddings.persist()
+        embeddings.persist(True)
         print(
             f"Added {len(added_tokens)} aliases: {list(zip(alias_placeholder_tokens, added_tokens, alias_initializer_tokens, added_ids))}"
         )
@@ -682,7 +682,7 @@ def main():
             f"Added {len(added_tokens)} tokens from embeddings dir: {list(zip(added_tokens, added_ids))}"
         )
 
-        embeddings.persist()
+        embeddings.persist(True)
 
     if len(args.placeholder_tokens) != 0:
         placeholder_token_ids, initializer_token_ids = add_placeholder_tokens(
diff --git a/train_lora.py b/train_lora.py
index 2a43252..eeac81f 100644
--- a/train_lora.py
+++ b/train_lora.py
@@ -777,7 +777,7 @@ def main():
             placeholder_tokens=alias_placeholder_tokens,
             initializer_tokens=alias_initializer_tokens,
         )
-        embeddings.persist()
+        embeddings.persist(True)
         print(
             f"Added {len(added_tokens)} aliases: {list(zip(alias_placeholder_tokens, added_tokens, alias_initializer_tokens, added_ids))}"
         )
@@ -806,7 +806,7 @@ def main():
         if args.train_dir_embeddings:
             print("Training embeddings from embeddings dir")
         else:
-            embeddings.persist()
+            embeddings.persist(True)
 
     if len(args.placeholder_tokens) != 0 and not args.train_dir_embeddings:
         embeddings = ensure_embeddings()
@@ -1117,7 +1117,7 @@ def main():
             no_val=True,
         )
 
-        embeddings.persist()
+        embeddings.persist(True)
 
     # LORA
     # --------------------------------------------------------------------------------
diff --git a/train_ti.py b/train_ti.py
index 89f4113..1d0cb6f 100644
--- a/train_ti.py
+++ b/train_ti.py
@@ -691,7 +691,7 @@ def main():
             placeholder_tokens=alias_placeholder_tokens,
             initializer_tokens=alias_initializer_tokens,
         )
-        embeddings.persist()
+        embeddings.persist(True)
         print(
             f"Added {len(added_tokens)} aliases: {list(zip(alias_placeholder_tokens, added_tokens, alias_initializer_tokens, added_ids))}"
         )
@@ -712,7 +712,7 @@ def main():
             args.placeholder_tokens = added_tokens
             print("Training embeddings from embeddings dir")
         else:
-            embeddings.persist()
+            embeddings.persist(True)
 
     if args.scale_lr:
         args.learning_rate = (
@@ -1067,7 +1067,7 @@ def main():
             args.train_data_template,
         ):
             run(i, [placeholder_token], [initializer_token], num_vectors, data_template)
-            embeddings.persist()
+            embeddings.persist(True)
 
 
 if __name__ == "__main__":
diff --git a/training/strategy/dreambooth.py b/training/strategy/dreambooth.py
index bd853e2..3d1abf7 100644
--- a/training/strategy/dreambooth.py
+++ b/training/strategy/dreambooth.py
@@ -98,7 +98,6 @@ def dreambooth_strategy_callbacks(
 
         if cycle < train_text_encoder_cycles:
             text_encoder.train()
-            tokenizer.train()
 
         yield
 
@@ -155,6 +154,8 @@ def dreambooth_strategy_callbacks(
         unet_ = accelerator.unwrap_model(unet, keep_fp32_wrapper=False)
         text_encoder_ = accelerator.unwrap_model(text_encoder, keep_fp32_wrapper=False)
 
+        text_encoder_.text_model.embeddings.persist(False)
+
         with ema_context():
             pipeline = VlpnStableDiffusion(
                 text_encoder=text_encoder_,