Add optional TI functionality to Dreambooth

3 files changed, 77 insertions, 45 deletions

diff --git a/dreambooth.py b/dreambooth.py
index da8399f..72c56cd 100644
--- a/dreambooth.py
+++ b/dreambooth.py
@@ -170,14 +170,14 @@ def parse_args():
     parser.add_argument(
         "--lr_warmup_steps",
         type=int,
-        default=300,
+        default=500,
         help="Number of steps for the warmup in the lr scheduler."
     )
     parser.add_argument(
         "--lr_cycles",
         type=int,
         default=None,
-        help="Number of restart cycles in the lr scheduler."
+        help="Number of restart cycles in the lr scheduler (if supported)."
     )
     parser.add_argument(
         "--use_ema",
@@ -506,11 +506,10 @@ def main():
 
     logging.basicConfig(filename=basepath.joinpath("log.txt"), level=logging.DEBUG)
 
-    save_args(basepath, args)
+    args.seed = args.seed or (torch.random.seed() >> 32)
+    set_seed(args.seed)
 
-    # If passed along, set the training seed now.
-    if args.seed is not None:
-        set_seed(args.seed)
+    save_args(basepath, args)
 
     # Load the tokenizer and add the placeholder token as a additional special token
     if args.tokenizer_name:
@@ -523,13 +522,22 @@ def main():
     vae = AutoencoderKL.from_pretrained(args.pretrained_model_name_or_path, subfolder='vae')
     unet = UNet2DConditionModel.from_pretrained(args.pretrained_model_name_or_path, subfolder='unet')
 
-    ema_unet = EMAModel(
-        unet,
-        inv_gamma=args.ema_inv_gamma,
-        power=args.ema_power,
-        max_value=args.ema_max_decay,
-        device=accelerator.device
-    ) if args.use_ema else None
+    ema_unet = None
+    if args.use_ema:
+        ema_unet = EMAModel(
+            unet,
+            inv_gamma=args.ema_inv_gamma,
+            power=args.ema_power,
+            max_value=args.ema_max_decay,
+            device=accelerator.device
+        )
+
+    if args.gradient_checkpointing:
+        unet.enable_gradient_checkpointing()
+        text_encoder.gradient_checkpointing_enable()
+
+    # Freeze text_encoder and vae
+    freeze_params(vae.parameters())
 
     if args.initializer_token is not None:
         # Convert the initializer_token, placeholder_token to ids
@@ -545,22 +553,22 @@ def main():
             print(f"Training new token {args.placeholder_token}.")
         placeholder_token_id = tokenizer.convert_tokens_to_ids(args.placeholder_token)
 
+        # Resize the token embeddings as we are adding new special tokens to the tokenizer
         text_encoder.resize_token_embeddings(len(tokenizer))
-        token_embeds = text_encoder.get_input_embeddings()
-        initializer_token_embeddings = token_embeds(initializer_token_ids)
-        token_embeds.weight.data[placeholder_token_id] = initializer_token_embeddings
-
-    prompt_processor = PromptProcessor(tokenizer, text_encoder)
 
-    if args.gradient_checkpointing:
-        unet.enable_gradient_checkpointing()
-        text_encoder.gradient_checkpointing_enable()
+        # Initialise the newly added placeholder token with the embeddings of the initializer token
+        token_embeds = text_encoder.get_input_embeddings().weight.data
+        original_token_embeds = token_embeds.detach().clone().to(accelerator.device)
+        initializer_token_embeddings = text_encoder.get_input_embeddings()(initializer_token_ids)
+        token_embeds[placeholder_token_id] = initializer_token_embeddings
 
-    # slice_size = unet.config.attention_head_dim // 2
-    # unet.set_attention_slice(slice_size)
+        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(),
+        ))
 
-    # Freeze text_encoder and vae
-    freeze_params(vae.parameters())
+    prompt_processor = PromptProcessor(tokenizer, text_encoder)
 
     if args.scale_lr:
         args.learning_rate_unet = (
@@ -583,6 +591,11 @@ def main():
     else:
         optimizer_class = torch.optim.AdamW
 
+    if args.initializer_token is not None:
+        text_encoder_params_to_optimize = text_encoder.get_input_embeddings().parameters()
+    else:
+        text_encoder_params_to_optimize = text_encoder.parameters()
+
     # Initialize the optimizer
     optimizer = optimizer_class(
         [
@@ -591,7 +604,7 @@ def main():
                 'lr': args.learning_rate_unet,
             },
             {
-                'params': text_encoder.parameters(),
+                'params': text_encoder_params_to_optimize,
                 'lr': args.learning_rate_text,
             }
         ],
@@ -849,9 +862,27 @@ def main():
                         loss = F.mse_loss(noise_pred.float(), noise.float(), reduction="mean")
 
                     accelerator.backward(loss)
+
+                    if args.initializer_token is not None:
+                        # Keep the token embeddings fixed except the newly added
+                        # embeddings for the concept, as we only want to optimize the concept embeddings
+                        if accelerator.num_processes > 1:
+                            token_embeds = text_encoder.module.get_input_embeddings().weight
+                        else:
+                            token_embeds = text_encoder.get_input_embeddings().weight
+
+                        # Get the index for tokens that we want to freeze
+                        index_fixed_tokens = torch.arange(len(tokenizer)) != placeholder_token_id
+                        token_embeds.data[index_fixed_tokens, :] = original_token_embeds[index_fixed_tokens, :]
+
                     if accelerator.sync_gradients:
-                        accelerator.clip_grad_norm_(itertools.chain(
-                            unet.parameters(), text_encoder.parameters()), args.max_grad_norm)
+                        params_to_clip = (
+                            unet.parameters()
+                            if args.initializer_token is not None
+                            else itertools.chain(unet.parameters(), text_encoder.parameters())
+                        )
+                        accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
+
                     optimizer.step()
                     if not accelerator.optimizer_step_was_skipped:
                         lr_scheduler.step()
@@ -896,8 +927,8 @@ def main():
             text_encoder.eval()
             val_loss = 0.0
 
-            for step, batch in enumerate(val_dataloader):
-                with torch.no_grad():
+            with torch.inference_mode():
+                for step, batch in enumerate(val_dataloader):
                     latents = vae.encode(batch["pixel_values"]).latent_dist.sample()
                     latents = latents * 0.18215
 
@@ -920,12 +951,12 @@ def main():
                     loss = loss.detach().item()
                     val_loss += loss
 
-                if accelerator.sync_gradients:
-                    local_progress_bar.update(1)
-                    global_progress_bar.update(1)
+                    if accelerator.sync_gradients:
+                        local_progress_bar.update(1)
+                        global_progress_bar.update(1)
 
-                logs = {"val/loss": loss}
-                local_progress_bar.set_postfix(**logs)
+                    logs = {"val/loss": loss}
+                    local_progress_bar.set_postfix(**logs)
 
             val_loss /= len(val_dataloader)
 
diff --git a/infer.py b/infer.py
index 8e17c4e..01010eb 100644
--- a/infer.py
+++ b/infer.py
@@ -258,7 +258,7 @@ def generate(output_dir, pipeline, args):
     output_dir = output_dir.joinpath(f"{now}_{slugify(args.prompt[0])[:100]}")
     output_dir.mkdir(parents=True, exist_ok=True)
 
-    seed = args.seed or torch.random.seed()
+    args.seed = args.seed or torch.random.seed()
 
     save_args(output_dir, args)
 
@@ -276,7 +276,7 @@ def generate(output_dir, pipeline, args):
                 dynamic_ncols=True
             )
 
-            generator = torch.Generator(device="cuda").manual_seed(seed + i)
+            generator = torch.Generator(device="cuda").manual_seed(args.seed + i)
             images = pipeline(
                 prompt=args.prompt * (args.batch_size // len(args.prompt)),
                 height=args.height,
@@ -290,7 +290,7 @@ def generate(output_dir, pipeline, args):
             ).images
 
             for j, image in enumerate(images):
-                image.save(output_dir.joinpath(f"{seed + i}_{j}.jpg"))
+                image.save(output_dir.joinpath(f"{args.seed + i}_{j}.jpg"))
 
     if torch.cuda.is_available():
         torch.cuda.empty_cache()
diff --git a/textual_inversion.py b/textual_inversion.py
index 8f266e0..fe56d36 100644
--- a/textual_inversion.py
+++ b/textual_inversion.py
@@ -521,6 +521,7 @@ def main():
 
     if args.gradient_checkpointing:
         unet.enable_gradient_checkpointing()
+        text_encoder.gradient_checkpointing_enable()
 
     # slice_size = unet.config.attention_head_dim // 2
     # unet.set_attention_slice(slice_size)
@@ -875,8 +876,8 @@ def main():
             text_encoder.eval()
             val_loss = 0.0
 
-            for step, batch in enumerate(val_dataloader):
-                with torch.no_grad():
+            with torch.inference_mode():
+                for step, batch in enumerate(val_dataloader):
                     latents = vae.encode(batch["pixel_values"]).latent_dist.sample()
                     latents = latents * 0.18215
 
@@ -899,12 +900,12 @@ def main():
                     loss = loss.detach().item()
                     val_loss += loss
 
-                if accelerator.sync_gradients:
-                    local_progress_bar.update(1)
-                    global_progress_bar.update(1)
+                    if accelerator.sync_gradients:
+                        local_progress_bar.update(1)
+                        global_progress_bar.update(1)
 
-                logs = {"val/loss": loss}
-                local_progress_bar.set_postfix(**logs)
+                    logs = {"val/loss": loss}
+                    local_progress_bar.set_postfix(**logs)
 
             val_loss /= len(val_dataloader)