Lora: Only register params with grad to optimizer

3 files changed, 7 insertions, 8 deletions

diff --git a/train_lora.py b/train_lora.py
index 8fc2d69..cf73645 100644
--- a/train_lora.py
+++ b/train_lora.py
@@ -662,9 +662,13 @@ def main():
         sample_frequency = math.ceil(num_train_epochs * (sample_frequency / args.num_train_steps))
 
     optimizer = create_optimizer(
-        itertools.chain(
-            unet.parameters(),
-            text_encoder.parameters(),
+        (
+            param
+            for param in itertools.chain(
+                unet.parameters(),
+                text_encoder.parameters(),
+            )
+            if param.requires_grad
         ),
         lr=args.learning_rate,
     )
diff --git a/training/strategy/lora.py b/training/strategy/lora.py
index 8905171..209785a 100644
--- a/training/strategy/lora.py
+++ b/training/strategy/lora.py
@@ -139,10 +139,6 @@ def lora_prepare(
     train_dataloader: DataLoader,
     val_dataloader: Optional[DataLoader],
     lr_scheduler: torch.optim.lr_scheduler._LRScheduler,
-    lora_rank: int = 4,
-    lora_alpha: int = 32,
-    lora_dropout: float = 0,
-    lora_bias: str = "none",
     **kwargs
 ):
     return accelerator.prepare(text_encoder, unet, optimizer, train_dataloader, val_dataloader, lr_scheduler) + ({},)
diff --git a/training/strategy/ti.py b/training/strategy/ti.py
index 677f5a3..c7520ed 100644
--- a/training/strategy/ti.py
+++ b/training/strategy/ti.py
@@ -209,7 +209,6 @@ def textual_inversion_prepare(
     text_encoder.text_model.final_layer_norm.requires_grad_(False)
     text_encoder.text_model.embeddings.position_embedding.requires_grad_(False)
     text_encoder.text_model.embeddings.token_embedding.requires_grad_(False)
-    text_encoder.eval()
 
     return text_encoder, unet, optimizer, train_dataloader, val_dataloader, lr_scheduler, {}