TI: Delta learning

1 files changed, 0 insertions, 23 deletions

diff --git a/training/strategy/ti.py b/training/strategy/ti.py
index c7520ed..16baa34 100644
--- a/training/strategy/ti.py
+++ b/training/strategy/ti.py
@@ -31,10 +31,6 @@ def textual_inversion_strategy_callbacks(
     seed: int,
     placeholder_tokens: list[str],
     placeholder_token_ids: list[list[int]],
-    gradient_checkpointing: bool = False,
-    use_emb_decay: bool = False,
-    emb_decay_target: float = 0.4,
-    emb_decay: float = 1e-2,
     use_ema: bool = False,
     ema_inv_gamma: float = 1.0,
     ema_power: int = 1,
@@ -106,28 +102,10 @@ def textual_inversion_strategy_callbacks(
             yield
 
     @torch.no_grad()
-    def on_before_optimize(lr: float, epoch: int):
-        if use_emb_decay:
-            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(zero_ids, lr: float):
         if ema_embeddings is not None:
             ema_embeddings.step(text_encoder.text_model.embeddings.temp_token_embedding.parameters())
 
-        if use_emb_decay:
-            lambda_ = emb_decay * lr
-
-            if lambda_ != 0:
-                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:
             return {"ema_decay": ema_embeddings.decay}
@@ -171,7 +149,6 @@ def textual_inversion_strategy_callbacks(
         on_accum_model=on_accum_model,
         on_train=on_train,
         on_eval=on_eval,
-        on_before_optimize=on_before_optimize,
         on_after_optimize=on_after_optimize,
         on_log=on_log,
         on_checkpoint=on_checkpoint,