1 files changed, 10 insertions, 8 deletions

diff --git a/training/strategy/ti.py b/training/strategy/ti.py
index 7ac5011..b9a5547 100644
--- a/training/strategy/ti.py
+++ b/training/strategy/ti.py
@@ -108,14 +108,11 @@ def textual_inversion_strategy_callbacks(
     @torch.no_grad()
     def on_before_optimize(lr: float, epoch: int):
         if use_emb_decay:
-            return torch.stack([
-                t
-                for t in text_encoder.text_model.embeddings.temp_token_embedding
-                if t.grad is not None
-            ])
+            w = text_encoder.text_model.embeddings.temp_token_embedding.weight
+            return torch.all(w.grad == 0, dim=1)
 
     @torch.no_grad()
-    def on_after_optimize(w, lr: float):
+    def on_after_optimize(zero_ids, lr: float):
         if ema_embeddings is not None:
             ema_embeddings.step(text_encoder.text_model.embeddings.temp_token_embedding.parameters())
 
@@ -123,8 +120,13 @@ def textual_inversion_strategy_callbacks(
             lambda_ = emb_decay * lr
 
             if lambda_ != 0:
-                norm = w[:, :].norm(dim=-1, keepdim=True)
-                w[:].add_((w[:] / norm.clamp_min(1e-12)) * lambda_ * (emb_decay_target - norm))
+                w = text_encoder.text_model.embeddings.temp_token_embedding.weight
+
+                mask = torch.ones(w.shape[0], dtype=torch.bool)
+                mask[zero_ids] = False
+
+                norm = w[mask, :].norm(dim=-1, keepdim=True)
+                w[mask].add_((w[mask] / norm.clamp_min(1e-12)) * lambda_ * (emb_decay_target - norm))
 
     def on_log():
         if ema_embeddings is not None: