Embedding normalization: Ignore tensors with grad = 0

1 files changed, 11 insertions, 4 deletions

diff --git a/training/strategy/ti.py b/training/strategy/ti.py
index 66d3129..09beec4 100644
--- a/training/strategy/ti.py
+++ b/training/strategy/ti.py
@@ -116,10 +116,17 @@ def textual_inversion_strategy_callbacks(
     @torch.no_grad()
     def on_before_optimize(lr: float, epoch: int):
         if use_emb_decay:
-            text_encoder.text_model.embeddings.normalize(
-                emb_decay_target,
-                min(1.0, emb_decay * lr)
-            )
+            lambda_ = emb_decay * lr
+
+            if lambda_ != 0:
+                w = text_encoder.text_model.embeddings.temp_token_embedding.weight
+
+                mask = torch.zeros(w.size(0), dtype=torch.bool)
+                mask[text_encoder.text_model.embeddings.temp_token_ids] = True
+                mask[torch.all(w.grad == 0, dim=1)] = 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_after_optimize(lr: float):
         if ema_embeddings is not None: