diff options
Diffstat (limited to 'training/strategy/ti.py')
| -rw-r--r-- | training/strategy/ti.py | 7 |
1 files changed, 4 insertions, 3 deletions
diff --git a/training/strategy/ti.py b/training/strategy/ti.py index 9df160a..55e9934 100644 --- a/training/strategy/ti.py +++ b/training/strategy/ti.py | |||
| @@ -107,18 +107,19 @@ def textual_inversion_strategy_callbacks( | |||
| 107 | @torch.no_grad() | 107 | @torch.no_grad() |
| 108 | def on_before_optimize(lr: float, epoch: int): | 108 | def on_before_optimize(lr: float, epoch: int): |
| 109 | if use_emb_decay: | 109 | if use_emb_decay: |
| 110 | return torch.stack([ | 110 | params = [ |
| 111 | p | 111 | p |
| 112 | for p in text_encoder.text_model.embeddings.token_override_embedding.params | 112 | for p in text_encoder.text_model.embeddings.token_override_embedding.params |
| 113 | if p.grad is not None | 113 | if p.grad is not None |
| 114 | ]) | 114 | ] |
| 115 | return torch.stack(params) if len(params) != 0 else None | ||
| 115 | 116 | ||
| 116 | @torch.no_grad() | 117 | @torch.no_grad() |
| 117 | def on_after_optimize(w, lr: float): | 118 | def on_after_optimize(w, lr: float): |
| 118 | if ema_embeddings is not None: | 119 | if ema_embeddings is not None: |
| 119 | ema_embeddings.step(text_encoder.text_model.embeddings.token_override_embedding.params.parameters()) | 120 | ema_embeddings.step(text_encoder.text_model.embeddings.token_override_embedding.params.parameters()) |
| 120 | 121 | ||
| 121 | if use_emb_decay: | 122 | if use_emb_decay and w is not None: |
| 122 | lambda_ = emb_decay * lr | 123 | lambda_ = emb_decay * lr |
| 123 | 124 | ||
| 124 | if lambda_ != 0: | 125 | if lambda_ != 0: |