Update

7 files changed, 72 insertions, 160 deletions

diff --git a/infer.py b/infer.py
index 8fdf63d..4648c0a 100644
--- a/infer.py
+++ b/infer.py
@@ -235,6 +235,7 @@ def load_embeddings(pipeline, embeddings_dir):
         pipeline.text_encoder.text_model.embeddings,
         Path(embeddings_dir)
     )
+    pipeline.text_encoder.text_model.embeddings.persist()
     print(f"Added {len(added_tokens)} tokens from embeddings dir: {list(zip(added_tokens, added_ids))}")
 
 
diff --git a/models/clip/embeddings.py b/models/clip/embeddings.py
index 6fda33c..dc4708a 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):
+    def __init__(self, config: CLIPTextConfig, embeddings: CLIPTextEmbeddings, dropout_p: float = 0.0):
         super().__init__(config)
 
         self.token_embedding = embeddings.token_embedding
@@ -46,6 +46,7 @@ class ManagedCLIPTextEmbeddings(CLIPTextEmbeddings):
 
         self.token_override_embedding = PseudoSparseEmbedding(
             self.token_embedding.embedding_dim,
+            dropout_p=dropout_p,
             device=self.token_embedding.weight.device,
             dtype=self.token_embedding.weight.dtype,
         )
@@ -134,7 +135,7 @@ class ManagedCLIPTextEmbeddings(CLIPTextEmbeddings):
         return embeddings
 
 
-def patch_managed_embeddings(text_encoder: CLIPTextModel) -> ManagedCLIPTextEmbeddings:
-    text_embeddings = ManagedCLIPTextEmbeddings(text_encoder.config, text_encoder.text_model.embeddings)
+def patch_managed_embeddings(text_encoder: CLIPTextModel, dropout_p: float = 0.0) -> ManagedCLIPTextEmbeddings:
+    text_embeddings = ManagedCLIPTextEmbeddings(text_encoder.config, text_encoder.text_model.embeddings, dropout_p)
     text_encoder.text_model.embeddings = text_embeddings
     return text_embeddings
diff --git a/models/sparse.py b/models/sparse.py
index d706db5..bcb2897 100644
--- a/models/sparse.py
+++ b/models/sparse.py
@@ -5,22 +5,29 @@ import torch.nn as nn
 
 
 class PseudoSparseEmbedding(nn.Module):
-    def __init__(self, embedding_dim: int, device=None, dtype=torch.float32):
+    def __init__(self, embedding_dim: int, dropout_p: float = 0.0, device=None, dtype=torch.float32):
         super().__init__()
 
         self.embedding_dim = embedding_dim
         self.dtype = dtype
         self.params = nn.ParameterList()
+
+        if dropout_p > 0.0:
+            self.dropout = nn.Dropout(p=dropout_p)
+        else:
+            self.dropout = lambda x: x
+
         self.register_buffer('mapping', torch.zeros(0, device=device, dtype=torch.long))
 
     def forward(self, input_ids: torch.LongTensor):
-        ids = self.mapping[input_ids.to(self.mapping.device)]
+        input_ids = input_ids.to(self.mapping.device)
+        ids = self.mapping[input_ids]
         mask = ~(ids == -1)
 
         if torch.all(~mask):
             embs = None
         else:
-            embs = torch.stack([self.params[id] for id in ids[mask]])
+            embs = self.dropout(torch.stack([self.params[id] for id in ids[mask]]))
 
         return embs, mask
 
diff --git a/train_dreambooth.py b/train_dreambooth.py
index f4d4cbb..2aca1e7 100644
--- a/train_dreambooth.py
+++ b/train_dreambooth.py
@@ -513,6 +513,7 @@ def main():
             raise ValueError("--embeddings_dir must point to an existing directory")
 
         added_tokens, added_ids = load_embeddings_from_dir(tokenizer, embeddings, embeddings_dir)
+        embeddings.persist()
         print(f"Added {len(added_tokens)} tokens from embeddings dir: {list(zip(added_tokens, added_ids))}")
 
     if args.scale_lr:
diff --git a/train_lora.py b/train_lora.py
index 8dbe45b..6e21634 100644
--- a/train_lora.py
+++ b/train_lora.py
@@ -159,12 +159,6 @@ def parse_args():
         help="Tag dropout probability.",
     )
     parser.add_argument(
-        "--pti_tag_dropout",
-        type=float,
-        default=0,
-        help="Tag dropout probability.",
-    )
-    parser.add_argument(
         "--no_tag_shuffle",
         action="store_true",
         help="Shuffle tags.",
@@ -236,28 +230,12 @@ def parse_args():
         default=2000
     )
     parser.add_argument(
-        "--num_pti_epochs",
-        type=int,
-        default=None
-    )
-    parser.add_argument(
-        "--num_pti_steps",
-        type=int,
-        default=500
-    )
-    parser.add_argument(
         "--gradient_accumulation_steps",
         type=int,
         default=1,
         help="Number of updates steps to accumulate before performing a backward/update pass.",
     )
     parser.add_argument(
-        "--pti_gradient_accumulation_steps",
-        type=int,
-        default=1,
-        help="Number of updates steps to accumulate before performing a backward/update pass.",
-    )
-    parser.add_argument(
         "--lora_r",
         type=int,
         default=8,
@@ -323,12 +301,6 @@ def parse_args():
         help="Initial learning rate (after the potential warmup period) to use.",
     )
     parser.add_argument(
-        "--learning_rate_pti",
-        type=float,
-        default=1e-4,
-        help="Initial learning rate (after the potential warmup period) to use.",
-    )
-    parser.add_argument(
         "--learning_rate_emb",
         type=float,
         default=1e-5,
@@ -467,12 +439,6 @@ def parse_args():
         help="How often to save a checkpoint and sample image",
     )
     parser.add_argument(
-        "--pti_sample_frequency",
-        type=int,
-        default=1,
-        help="How often to save a checkpoint and sample image",
-    )
-    parser.add_argument(
         "--sample_image_size",
         type=int,
         default=768,
@@ -509,12 +475,6 @@ def parse_args():
         help="Batch size (per device) for the training dataloader."
     )
     parser.add_argument(
-        "--pti_batch_size",
-        type=int,
-        default=1,
-        help="Batch size (per device) for the training dataloader."
-    )
-    parser.add_argument(
         "--sample_steps",
         type=int,
         default=10,
@@ -527,6 +487,12 @@ def parse_args():
         help="The weight of prior preservation loss."
     )
     parser.add_argument(
+        "--emb_dropout",
+        type=float,
+        default=0,
+        help="Embedding dropout probability.",
+    )
+    parser.add_argument(
         "--use_emb_decay",
         action="store_true",
         help="Whether to use embedding decay."
@@ -674,7 +640,7 @@ 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.pretrained_model_name_or_path, args.emb_dropout)
 
     unet_config = LoraConfig(
         r=args.lora_r,
@@ -720,6 +686,7 @@ def main():
             raise ValueError("--embeddings_dir must point to an existing directory")
 
         added_tokens, added_ids = load_embeddings_from_dir(tokenizer, embeddings, embeddings_dir)
+        embeddings.persist()
         print(f"Added {len(added_tokens)} tokens from embeddings dir: {list(zip(added_tokens, added_ids))}")
 
     placeholder_token_ids, initializer_token_ids = add_placeholder_tokens(
@@ -744,19 +711,14 @@ def main():
             args.learning_rate_text * args.gradient_accumulation_steps *
             args.train_batch_size * accelerator.num_processes
         )
-        args.learning_rate_pti = (
-            args.learning_rate_pti * args.pti_gradient_accumulation_steps *
-            args.pti_batch_size * accelerator.num_processes
-        )
         args.learning_rate_emb = (
-            args.learning_rate_emb * args.pti_gradient_accumulation_steps *
-            args.pti_batch_size * accelerator.num_processes
+            args.learning_rate_emb * args.gradient_accumulation_steps *
+            args.train_batch_size * accelerator.num_processes
         )
 
     if args.find_lr:
         args.learning_rate_unet = 1e-6
         args.learning_rate_text = 1e-6
-        args.learning_rate_pti = 1e-6
         args.learning_rate_emb = 1e-6
         args.lr_scheduler = "exponential_growth"
 
@@ -817,7 +779,6 @@ def main():
         args.lr_min_lr = args.learning_rate_unet
         args.learning_rate_unet = None
         args.learning_rate_text = None
-        args.learning_rate_pti = None
         args.learning_rate_emb = None
     elif args.optimizer == 'dadam':
         try:
@@ -836,7 +797,6 @@ def main():
 
         args.learning_rate_unet = 1.0
         args.learning_rate_text = 1.0
-        args.learning_rate_pti = 1.0
         args.learning_rate_emb = 1.0
     elif args.optimizer == 'dadan':
         try:
@@ -853,7 +813,6 @@ def main():
 
         args.learning_rate_unet = 1.0
         args.learning_rate_text = 1.0
-        args.learning_rate_pti = 1.0
         args.learning_rate_emb = 1.0
     else:
         raise ValueError(f"Unknown --optimizer \"{args.optimizer}\"")
@@ -920,80 +879,6 @@ def main():
         mid_point=args.lr_mid_point,
     )
 
-    # PTI
-    # --------------------------------------------------------------------------------
-
-    if len(args.placeholder_tokens) != 0:
-        pti_datamodule = create_datamodule(
-            batch_size=args.pti_batch_size,
-            dropout=args.pti_tag_dropout,
-            filter=partial(keyword_filter, args.filter_tokens, args.collection, args.exclude_collections),
-        )
-        pti_datamodule.setup()
-
-        num_pti_epochs = args.num_pti_epochs
-        pti_sample_frequency = args.pti_sample_frequency
-        if num_pti_epochs is None:
-            num_pti_epochs = math.ceil(
-                args.num_pti_steps / len(pti_datamodule.train_dataset)
-            ) * args.pti_gradient_accumulation_steps
-            pti_sample_frequency = math.ceil(num_pti_epochs * (pti_sample_frequency / args.num_pti_steps))
-
-        if num_pti_epochs > 0:
-            pti_optimizer = create_optimizer(
-                [
-                    {
-                        "params": text_encoder.text_model.embeddings.token_override_embedding.parameters(),
-                        "lr": args.learning_rate_pti,
-                        "weight_decay": 0,
-                    },
-                ]
-            )
-
-            pti_lr_scheduler = create_lr_scheduler(
-                gradient_accumulation_steps=args.pti_gradient_accumulation_steps,
-                optimizer=pti_optimizer,
-                num_training_steps_per_epoch=len(pti_datamodule.train_dataloader),
-                train_epochs=num_pti_epochs,
-            )
-
-            continue_training = True
-            training_iter = 1
-
-            while continue_training:
-                print("")
-                print(f"============ PTI cycle {training_iter} ============")
-                print("")
-
-                pti_project = f"pti_{training_iter}"
-                pti_output_dir = output_dir / pti_project
-                pti_checkpoint_output_dir = pti_output_dir / "model"
-                pti_sample_output_dir = pti_output_dir / "samples"
-
-                trainer(
-                    strategy=lora_strategy,
-                    pti_mode=True,
-                    project=pti_project,
-                    train_dataloader=pti_datamodule.train_dataloader,
-                    val_dataloader=pti_datamodule.val_dataloader,
-                    optimizer=pti_optimizer,
-                    lr_scheduler=pti_lr_scheduler,
-                    num_train_epochs=num_pti_epochs,
-                    gradient_accumulation_steps=args.pti_gradient_accumulation_steps,
-                    # --
-                    group_labels=["emb"],
-                    sample_output_dir=pti_sample_output_dir,
-                    checkpoint_output_dir=pti_checkpoint_output_dir,
-                    sample_frequency=pti_sample_frequency,
-                )
-
-                response = input("Run another cycle? [y/n] ")
-                continue_training = response.lower().strip() != "n"
-                training_iter += 1
-
-            if not args.train_emb:
-                embeddings.persist()
-
     # LORA
     # --------------------------------------------------------------------------------
 
@@ -1062,9 +947,8 @@ def main():
         print("")
 
         lora_project = f"lora_{training_iter}"
-        lora_output_dir = output_dir / lora_project
-        lora_checkpoint_output_dir = lora_output_dir / "model"
-        lora_sample_output_dir = lora_output_dir / "samples"
+        lora_checkpoint_output_dir = output_dir / lora_project / "model"
+        lora_sample_output_dir = output_dir / lora_project / "samples"
 
         trainer(
             strategy=lora_strategy,
diff --git a/train_ti.py b/train_ti.py
index daf8bc5..2d51800 100644
--- a/train_ti.py
+++ b/train_ti.py
@@ -458,6 +458,12 @@ def parse_args():
         help="The weight of prior preservation loss."
     )
     parser.add_argument(
+        "--emb_dropout",
+        type=float,
+        default=0,
+        help="Embedding dropout probability.",
+    )
+    parser.add_argument(
         "--use_emb_decay",
         action="store_true",
         help="Whether to use embedding decay."
@@ -624,7 +630,7 @@ 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.pretrained_model_name_or_path, args.emb_dropout)
 
     tokenizer.set_use_vector_shuffle(args.vector_shuffle)
     tokenizer.set_dropout(args.vector_dropout)
@@ -755,8 +761,6 @@ def main():
     else:
         raise ValueError(f"Unknown --optimizer \"{args.optimizer}\"")
 
-    checkpoint_output_dir = output_dir / "checkpoints"
-
     trainer = partial(
         train,
         accelerator=accelerator,
@@ -777,7 +781,6 @@ def main():
         global_step_offset=global_step_offset,
         offset_noise_strength=args.offset_noise_strength,
         # --
-        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,
@@ -793,11 +796,6 @@ def main():
     )
 
     def run(i: int, placeholder_tokens: list[str], initializer_tokens: list[str], num_vectors: Union[int, list[int]], data_template: str):
-        if len(placeholder_tokens) == 1:
-            sample_output_dir = output_dir / f"samples_{placeholder_tokens[0]}"
-        else:
-            sample_output_dir = output_dir / "samples"
-
         placeholder_token_ids, initializer_token_ids = add_placeholder_tokens(
             tokenizer=tokenizer,
             embeddings=embeddings,
@@ -809,7 +807,11 @@ def main():
 
         stats = list(zip(placeholder_tokens, placeholder_token_ids, initializer_tokens, initializer_token_ids))
 
-        print(f"{i + 1}: {stats}")
+        print("")
+        print(f"============ TI batch {i + 1} ============")
+        print("")
+        print(stats)
+        print("")
 
         filter_tokens = [token for token in args.filter_tokens if token in placeholder_tokens]
 
@@ -868,20 +870,36 @@ def main():
             mid_point=args.lr_mid_point,
         )
 
-        trainer(
-            project="textual_inversion",
-            train_dataloader=datamodule.train_dataloader,
-            val_dataloader=datamodule.val_dataloader,
-            optimizer=optimizer,
-            lr_scheduler=lr_scheduler,
-            num_train_epochs=num_train_epochs,
-            # --
-            group_labels=["emb"],
-            sample_output_dir=sample_output_dir,
-            sample_frequency=sample_frequency,
-            placeholder_tokens=placeholder_tokens,
-            placeholder_token_ids=placeholder_token_ids,
-        )
+        continue_training = True
+        training_iter = 1
+
+        while continue_training:
+            print(f"------------ TI cycle {training_iter} ------------")
+            print("")
+
+            project = f"{placeholder_tokens[0]}_{training_iter}" if len(placeholder_tokens) == 1 else f"{training_iter}"
+            sample_output_dir = output_dir / project / "samples"
+            checkpoint_output_dir = output_dir / project / "checkpoints"
+
+            trainer(
+                project=project,
+                train_dataloader=datamodule.train_dataloader,
+                val_dataloader=datamodule.val_dataloader,
+                optimizer=optimizer,
+                lr_scheduler=lr_scheduler,
+                num_train_epochs=num_train_epochs,
+                # --
+                group_labels=["emb"],
+                checkpoint_output_dir=checkpoint_output_dir,
+                sample_output_dir=sample_output_dir,
+                sample_frequency=sample_frequency,
+                placeholder_tokens=placeholder_tokens,
+                placeholder_token_ids=placeholder_token_ids,
+            )
+
+            response = input("Run another cycle? [y/n] ")
+            continue_training = response.lower().strip() != "n"
+            training_iter += 1
 
     if not args.sequential:
         run(0, args.placeholder_tokens, args.initializer_tokens, args.num_vectors, args.train_data_template)
diff --git a/training/functional.py b/training/functional.py
index 7d49782..e14aeea 100644
--- a/training/functional.py
+++ b/training/functional.py
@@ -72,7 +72,7 @@ def make_grid(images, rows, cols):
     return grid
 
 
-def get_models(pretrained_model_name_or_path: str):
+def get_models(pretrained_model_name_or_path: str, emb_dropout: float = 0.0):
     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')
@@ -81,7 +81,7 @@ def get_models(pretrained_model_name_or_path: str):
     sample_scheduler = UniPCMultistepScheduler.from_pretrained(
         pretrained_model_name_or_path, subfolder='scheduler')
 
-    embeddings = patch_managed_embeddings(text_encoder)
+    embeddings = patch_managed_embeddings(text_encoder, emb_dropout)
 
     return tokenizer, text_encoder, vae, unet, noise_scheduler, sample_scheduler, embeddings