Improved Textual Inversion: Completely exclude untrained embeddings from training

2 files changed, 76 insertions, 18 deletions

diff --git a/train_ti.py b/train_ti.py
index 198cf37..bb51dc2 100644
--- a/train_ti.py
+++ b/train_ti.py
@@ -24,7 +24,8 @@ from common import load_text_embeddings, load_text_embedding
 from pipelines.stable_diffusion.vlpn_stable_diffusion import VlpnStableDiffusion
 from data.csv import CSVDataModule, CSVDataItem
 from training.optimization import get_one_cycle_schedule
-from training.util import AverageMeter, CheckpointerBase, freeze_params, save_args
+from training.ti import patch_trainable_embeddings
+from training.util import AverageMeter, CheckpointerBase, save_args
 from models.clip.prompt import PromptProcessor
 
 logger = get_logger(__name__)
@@ -512,24 +513,14 @@ def main():
         for (token_id, token) in zip(placeholder_token_id, args.placeholder_token):
             load_text_embedding(token_embeds, token_id, resumepath.joinpath(f"{token}_{args.global_step}_end.bin"))
 
-    original_token_embeds = token_embeds.clone().to(accelerator.device)
-
     initializer_token_embeddings = text_encoder.get_input_embeddings()(initializer_token_ids)
     for (token_id, embeddings) in zip(placeholder_token_id, initializer_token_embeddings):
         token_embeds[token_id] = embeddings
 
-    index_fixed_tokens = torch.arange(len(tokenizer))
-    index_fixed_tokens = index_fixed_tokens[~torch.isin(index_fixed_tokens, torch.tensor(placeholder_token_id))]
+    vae.requires_grad_(False)
+    unet.requires_grad_(False)
 
-    # Freeze vae and unet
-    freeze_params(vae.parameters())
-    freeze_params(unet.parameters())
-    # Freeze all parameters except for the token embeddings in text encoder
-    freeze_params(itertools.chain(
-        text_encoder.text_model.encoder.parameters(),
-        text_encoder.text_model.final_layer_norm.parameters(),
-        text_encoder.text_model.embeddings.position_embedding.parameters(),
-    ))
+    text_embeddings = patch_trainable_embeddings(text_encoder, placeholder_token_id)
 
     prompt_processor = PromptProcessor(tokenizer, text_encoder)
 
@@ -843,10 +834,7 @@ def main():
                         lr_scheduler.step()
                     optimizer.zero_grad(set_to_none=True)
 
-                    # Let's make sure we don't update any embedding weights besides the newly added token
-                    with torch.no_grad():
-                        text_encoder.get_input_embeddings(
-                        ).weight[index_fixed_tokens] = original_token_embeds[index_fixed_tokens]
+                    text_embeddings.save()
 
                     avg_loss.update(loss.detach_(), bsz)
                     avg_acc.update(acc.detach_(), bsz)
diff --git a/training/ti.py b/training/ti.py
new file mode 100644
index 0000000..a5fd8e4
--- /dev/null
+++ b/training/ti.py
@@ -0,0 +1,70 @@
+from typing import Optional
+
+import torch
+import torch.nn as nn
+
+from transformers.models.clip import CLIPTextModel, CLIPTextConfig
+from transformers.models.clip.modeling_clip import CLIPTextEmbeddings
+
+
+def patch_trainable_embeddings(text_encoder: CLIPTextModel, new_ids: list[int]):
+    text_embeddings = TrainableEmbeddings(text_encoder.config, new_ids)
+    text_embeddings.token_embedding.weight = text_encoder.text_model.embeddings.token_embedding.weight
+    text_embeddings.position_embedding.weight = text_encoder.text_model.embeddings.position_embedding.weight
+    text_encoder.text_model.embeddings = text_embeddings
+    return text_embeddings
+
+
+class TrainableEmbeddings(CLIPTextEmbeddings):
+    def __init__(self, config: CLIPTextConfig, new_ids: list[int]):
+        super().__init__(config)
+
+        self.token_embedding.requires_grad_(False)
+        self.position_embedding.requires_grad_(False)
+
+        self.id_mapping = {new_ids[i]: i for i in range(len(new_ids))}
+
+        indices = torch.arange(self.token_embedding.num_embeddings)
+        self.train_indices = indices[torch.isin(indices, torch.tensor(new_ids))]
+
+        self.trainable_embedding = nn.Embedding.from_pretrained(self.token_embedding.weight[self.train_indices])
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+    ) -> torch.Tensor:
+        seq_length = input_ids.shape[-1] if input_ids is not None else inputs_embeds.shape[-2]
+
+        if position_ids is None:
+            position_ids = self.position_ids[:, :seq_length]
+
+        if inputs_embeds is None:
+            mask = torch.isin(
+                input_ids,
+                self.train_indices.to(input_ids.device)
+            ).unsqueeze(-1).expand(-1, -1, self.token_embedding.embedding_dim)
+
+            trainable_input_ids = torch.tensor([
+                [
+                    self.id_mapping[id] if id in self.id_mapping else 0
+                    for id in batch
+                ]
+                for batch in input_ids
+            ], device=input_ids.device)
+
+            inputs_embeds = torch.where(
+                mask,
+                self.trainable_embedding(trainable_input_ids),
+                self.token_embedding(input_ids)
+            )
+
+        position_embeddings = self.position_embedding(position_ids)
+        embeddings = inputs_embeds + position_embeddings
+
+        return embeddings
+
+    @torch.no_grad()
+    def save(self):
+        self.token_embedding.weight.data[self.train_indices] = self.trainable_embedding.weight.data