1 files changed, 5 insertions, 4 deletions
diff --git a/training/strategy/ti.py b/training/strategy/ti.py
index 1af834b..ba78b98 100644
--- a/training/strategy/ti.py
+++ b/training/strategy/ti.py
@@ -119,10 +119,11 @@ def textual_inversion_strategy_callbacks(
    def on_after_optimize(lr: float):
        if use_emb_decay:
-            text_encoder.text_model.embeddings.normalize(
+            with torch.no_grad():
-                emb_decay_target,
+                text_encoder.text_model.embeddings.normalize(
-                min(1.0, emb_decay * lr)
+                    emb_decay_target,
-            )
+                    min(1.0, emb_decay * lr)
+                )
        if ema_embeddings is not None:
            ema_embeddings.step(text_encoder.text_model.embeddings.temp_token_embedding.parameters())

diff --git a/training/strategy/ti.py b/training/strategy/ti.py index 1af834b..ba78b98 100644 --- a/training/strategy/ti.py +++ b/training/strategy/ti.py
@@ -119,10 +119,11 @@ def textual_inversion_strategy_callbacks(
119		119
120	def on_after_optimize(lr: float):	120	def on_after_optimize(lr: float):
121	if use_emb_decay:	121	if use_emb_decay:
122	text_encoder.text_model.embeddings.normalize(	122	with torch.no_grad():
123	emb_decay_target,	123	text_encoder.text_model.embeddings.normalize(
124	min(1.0, emb_decay * lr)	124	emb_decay_target,
125	)	125	min(1.0, emb_decay * lr)
		126	)
126		127
127	if ema_embeddings is not None:	128	if ema_embeddings is not None:
128	ema_embeddings.step(text_encoder.text_model.embeddings.temp_token_embedding.parameters())	129	ema_embeddings.step(text_encoder.text_model.embeddings.temp_token_embedding.parameters())