TI: Bring back old embedding decay

5 files changed, 57 insertions, 22 deletions

diff --git a/models/clip/embeddings.py b/models/clip/embeddings.py
index a356434..63a141f 100644
--- a/models/clip/embeddings.py
+++ b/models/clip/embeddings.py
@@ -37,7 +37,7 @@ def resize_embedding(old_embedding: nn.Embedding, new_num_embeddings: int, initi
 
 
 class ManagedCLIPTextEmbeddings(CLIPTextEmbeddings):
-    def __init__(self, config: CLIPTextConfig, embeddings: CLIPTextEmbeddings, alpha: float = 1.0):
+    def __init__(self, config: CLIPTextConfig, embeddings: CLIPTextEmbeddings):
         super().__init__(config)
 
         self.token_embedding = embeddings.token_embedding
@@ -49,7 +49,6 @@ class ManagedCLIPTextEmbeddings(CLIPTextEmbeddings):
             device=self.token_embedding.weight.device,
             dtype=self.token_embedding.weight.dtype,
         )
-        self.alpha = alpha
 
     def resize(self, size: int):
         self.token_override_embedding.resize(size)
@@ -87,7 +86,7 @@ class ManagedCLIPTextEmbeddings(CLIPTextEmbeddings):
         token_ids = torch.tensor(token_ids, dtype=torch.long)
 
         self.token_embedding.weight.data[token_ids] = initializer
-        self.token_override_embedding.set(token_ids)
+        self.token_override_embedding.set(token_ids, initializer)
 
     def load_embed(self, input_ids: list[int], filename: Path):
         with safe_open(filename, framework="pt", device="cpu") as file:
@@ -101,8 +100,8 @@ class ManagedCLIPTextEmbeddings(CLIPTextEmbeddings):
         embs, mask = self.token_override_embedding(input_ids)
         if embs is not None:
             input_ids = input_ids[mask]
-            self.token_embedding.weight.data[input_ids] += self.alpha * embs
-        self.token_override_embedding.unset(input_ids)
+            self.token_embedding.weight.data[input_ids] = embs
+            self.token_override_embedding.unset(input_ids)
 
     def get_embed(self, input_ids: Union[list[int], torch.LongTensor]):
         if isinstance(input_ids, list):
@@ -111,7 +110,7 @@ class ManagedCLIPTextEmbeddings(CLIPTextEmbeddings):
         embs = self.token_embedding(input_ids)
         embs_override, mask = self.token_override_embedding(input_ids)
         if embs_override is not None:
-            embs[mask] += self.alpha * embs_override
+            embs[mask] = embs_override
 
         return embs
 
@@ -135,7 +134,7 @@ class ManagedCLIPTextEmbeddings(CLIPTextEmbeddings):
         return embeddings
 
 
-def patch_managed_embeddings(text_encoder: CLIPTextModel, alpha: float = 1.0) -> ManagedCLIPTextEmbeddings:
-    text_embeddings = ManagedCLIPTextEmbeddings(text_encoder.config, text_encoder.text_model.embeddings, alpha)
+def patch_managed_embeddings(text_encoder: CLIPTextModel) -> ManagedCLIPTextEmbeddings:
+    text_embeddings = ManagedCLIPTextEmbeddings(text_encoder.config, text_encoder.text_model.embeddings)
     text_encoder.text_model.embeddings = text_embeddings
     return text_embeddings
diff --git a/models/sparse.py b/models/sparse.py
index 0b15454..8910316 100644
--- a/models/sparse.py
+++ b/models/sparse.py
@@ -13,10 +13,7 @@ class PseudoSparseEmbedding(nn.Module):
         self.params = nn.ParameterList()
         self.mapping = torch.zeros(0, device=device, dtype=torch.long)
 
-    def forward(self, input_ids: Optional[torch.LongTensor] = None):
-        if input_ids is None:
-            input_ids = torch.arange(self.mapping.shape[0])
-
+    def forward(self, input_ids: torch.LongTensor):
         ids = self.mapping[input_ids.to(self.mapping.device)]
         mask = ~(ids == -1)
 
@@ -43,6 +40,12 @@ class PseudoSparseEmbedding(nn.Module):
             else:
                 return [self.set(id) for id in input_ids]
 
+        if tensor is None:
+            tensor = torch.zeros(self.embedding_dim, device=self.mapping.device, dtype=self.dtype)
+
+        if tensor.shape[-1] != self.embedding_dim:
+            raise ValueError(f"Expected tensor of shape [..., {self.embedding_dim}], but got [..., {tensor.shape[-1]}]")
+
         id = self.mapping[input_ids]
 
         if id == -1:
@@ -50,8 +53,7 @@ class PseudoSparseEmbedding(nn.Module):
             self.mapping[input_ids] = id
             self.params.append(torch.zeros(self.embedding_dim, device=self.mapping.device, dtype=self.dtype))
 
-        self.params[id] = tensor if tensor is not None else torch.zeros(
-            self.embedding_dim, device=self.mapping.device, dtype=self.dtype)
+        self.params[id] = tensor
 
     def unset(self, input_ids: torch.LongTensor):
         self.mapping[input_ids] = -1
diff --git a/train_ti.py b/train_ti.py
index a9a2333..4366c9e 100644
--- a/train_ti.py
+++ b/train_ti.py
@@ -353,7 +353,7 @@ def parse_args():
     parser.add_argument(
         "--adam_weight_decay",
         type=float,
-        default=1e-2,
+        default=0,
         help="Weight decay to use."
     )
     parser.add_argument(
@@ -451,10 +451,21 @@ def parse_args():
         help="The weight of prior preservation loss."
     )
     parser.add_argument(
-        "--emb_alpha",
-        default=1.0,
+        "--use_emb_decay",
+        action="store_true",
+        help="Whether to use embedding decay."
+    )
+    parser.add_argument(
+        "--emb_decay_target",
+        default=0.4,
+        type=float,
+        help="Embedding decay target."
+    )
+    parser.add_argument(
+        "--emb_decay",
+        default=1e+2,
         type=float,
-        help="Embedding alpha."
+        help="Embedding decay factor."
     )
     parser.add_argument(
         "--noise_timesteps",
@@ -600,7 +611,7 @@ def main():
     save_args(output_dir, args)
 
     tokenizer, text_encoder, vae, unet, noise_scheduler, sample_scheduler, embeddings = get_models(
-        args.pretrained_model_name_or_path, args.emb_alpha)
+        args.pretrained_model_name_or_path)
 
     tokenizer.set_use_vector_shuffle(args.vector_shuffle)
     tokenizer.set_dropout(args.vector_dropout)
@@ -744,6 +755,9 @@ def main():
         tokenizer=tokenizer,
         sample_scheduler=sample_scheduler,
         checkpoint_output_dir=checkpoint_output_dir,
+        use_emb_decay=args.use_emb_decay,
+        emb_decay_target=args.emb_decay_target,
+        emb_decay=args.emb_decay,
         use_ema=args.use_ema,
         ema_inv_gamma=args.ema_inv_gamma,
         ema_power=args.ema_power,
diff --git a/training/functional.py b/training/functional.py
index 1d8e2ee..96ecbc1 100644
--- a/training/functional.py
+++ b/training/functional.py
@@ -73,7 +73,7 @@ def make_grid(images, rows, cols):
     return grid
 
 
-def get_models(pretrained_model_name_or_path: str, emb_alpha: float = 1.0):
+def get_models(pretrained_model_name_or_path: str):
     tokenizer = MultiCLIPTokenizer.from_pretrained(pretrained_model_name_or_path, subfolder='tokenizer')
     text_encoder = CLIPTextModel.from_pretrained(pretrained_model_name_or_path, subfolder='text_encoder')
     vae = AutoencoderKL.from_pretrained(pretrained_model_name_or_path, subfolder='vae')
@@ -82,7 +82,7 @@ def get_models(pretrained_model_name_or_path: str, emb_alpha: float = 1.0):
     sample_scheduler = UniPCMultistepScheduler.from_pretrained(
         pretrained_model_name_or_path, subfolder='scheduler')
 
-    embeddings = patch_managed_embeddings(text_encoder, emb_alpha)
+    embeddings = patch_managed_embeddings(text_encoder)
 
     return tokenizer, text_encoder, vae, unet, noise_scheduler, sample_scheduler, embeddings
 
diff --git a/training/strategy/ti.py b/training/strategy/ti.py
index 95128da..9df160a 100644
--- a/training/strategy/ti.py
+++ b/training/strategy/ti.py
@@ -31,6 +31,9 @@ def textual_inversion_strategy_callbacks(
     seed: int,
     placeholder_tokens: list[str],
     placeholder_token_ids: list[list[int]],
+    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,
@@ -102,10 +105,26 @@ def textual_inversion_strategy_callbacks(
             yield
 
     @torch.no_grad()
-    def on_after_optimize(zero_ids, lr: float):
+    def on_before_optimize(lr: float, epoch: int):
+        if use_emb_decay:
+            return torch.stack([
+                p
+                for p in text_encoder.text_model.embeddings.token_override_embedding.params
+                if p.grad is not None
+            ])
+
+    @torch.no_grad()
+    def on_after_optimize(w, lr: float):
         if ema_embeddings is not None:
             ema_embeddings.step(text_encoder.text_model.embeddings.token_override_embedding.params.parameters())
 
+        if use_emb_decay:
+            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))
+
     def on_log():
         if ema_embeddings is not None:
             return {"ema_decay": ema_embeddings.decay}
@@ -149,6 +168,7 @@ 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,