Merge pull request #130 from cloneofsimo/develop

v0.1.1

lora_diffusion/cli_lora_pti.py

@@ -417,6 +417,7 @@ def perform_tuning(
     placeholder_token_ids,
     placeholder_tokens,
     save_path,
+    lr_scheduler_lora,
 ):
 
     progress_bar = tqdm(range(num_steps))
@@ -430,6 +431,8 @@ def perform_tuning(
 
     for epoch in range(math.ceil(num_steps / len(dataloader))):
         for batch in dataloader:
+            lr_scheduler_lora.step()
+
             optimizer.zero_grad()
 
             loss = loss_step(
@@ -447,6 +450,11 @@ def perform_tuning(
             )
             optimizer.step()
             progress_bar.update(1)
+            logs = {
+                "loss": loss.detach().item(),
+                "lr": lr_scheduler_lora.get_last_lr()[0],
+            }
+            progress_bar.set_postfix(**logs)
 
             global_step += 1
 
@@ -504,27 +512,29 @@ def train(
     color_jitter: bool = True,
     train_batch_size: int = 1,
     sample_batch_size: int = 1,
-    max_train_steps_tuning: int = 10000,
-    max_train_steps_ti: int = 2000,
-    save_steps: int = 500,
-    gradient_accumulation_steps: int = 1,
+    max_train_steps_tuning: int = 1000,
+    max_train_steps_ti: int = 1000,
+    save_steps: int = 100,
+    gradient_accumulation_steps: int = 4,
     gradient_checkpointing: bool = False,
     mixed_precision="fp16",
     lora_rank: int = 4,
     lora_unet_target_modules={"CrossAttention", "Attention", "GEGLU"},
     lora_clip_target_modules={"CLIPAttention"},
     clip_ti_decay: bool = True,
-    learning_rate_unet: float = 1e-5,
+    learning_rate_unet: float = 1e-4,
     learning_rate_text: float = 1e-5,
     learning_rate_ti: float = 5e-4,
     continue_inversion: bool = True,
     continue_inversion_lr: Optional[float] = None,
     use_face_segmentation_condition: bool = False,
     scale_lr: bool = False,
-    lr_scheduler: str = "constant",
+    lr_scheduler: str = "linear",
     lr_warmup_steps: int = 0,
-    weight_decay_ti: float = 0.01,
-    weight_decay_lora: float = 0.01,
+    lr_scheduler_lora: str = "linear",
+    lr_warmup_steps_lora: int = 0,
+    weight_decay_ti: float = 0.00,
+    weight_decay_lora: float = 0.001,
     use_8bit_adam: bool = False,
     device="cuda:0",
     extra_args: Optional[dict] = None,
@@ -553,7 +563,7 @@ def train(
     placeholder_tokens = placeholder_tokens.split("|")
     if initializer_tokens is None:
         print("PTI : Initializer Token not give, random inits")
-        initializer_tokens = ["<rand-0.036>"] * len(placeholder_tokens)
+        initializer_tokens = ["<rand-0.017>"] * len(placeholder_tokens)
     else:
         initializer_tokens = initializer_tokens.split("|")
 
@@ -588,8 +598,7 @@ def train(
     )
 
     if gradient_checkpointing:
-        text_encoder.gradient_checkpointing_enable()
-        unet.gradient_checkpointing_enable()
+        unet.enable_gradient_checkpointing()
 
     if scale_lr:
         unet_lr = learning_rate_unet * gradient_accumulation_steps * train_batch_size
@@ -734,6 +743,13 @@ def train(
 
     train_dataset.blur_amount = 70
 
+    lr_scheduler_lora = get_scheduler(
+        lr_scheduler_lora,
+        optimizer=lora_optimizers,
+        num_warmup_steps=lr_warmup_steps_lora,
+        num_training_steps=max_train_steps_tuning,
+    )
+
     perform_tuning(
         unet,
         vae,
@@ -746,6 +762,7 @@ def train(
         placeholder_tokens=placeholder_tokens,
         placeholder_token_ids=placeholder_token_ids,
         save_path=output_dir,
+        lr_scheduler_lora=lr_scheduler_lora,
     )
 
 

lora_diffusion/utils.py

@@ -51,16 +51,15 @@
 ]
 
 
-def image_grid(_imgs, rows = None, cols = None):
-    
+def image_grid(_imgs, rows=None, cols=None):
+
     if rows is None and cols is None:
         rows = cols = math.ceil(len(_imgs) ** 0.5)
-    
+
     if rows is None:
         rows = math.ceil(len(_imgs) / cols)
     if cols is None:
         cols = math.ceil(len(_imgs) / rows)
-
 
     w, h = _imgs[0].size
     grid = Image.new("RGB", size=(cols * w, rows * h))
@@ -176,25 +175,23 @@ def visualize_progress(
     text_sclae=1.0,
     num_inference_steps=50,
     guidance_scale=5.0,
-    offset : int = 0,
-    limit : int = 10,
-    seed : int = 0
+    offset: int = 0,
+    limit: int = 10,
+    seed: int = 0,
 ):
 
-
     imgs = []
     if isinstance(path_alls, str):
         alls = list(set(glob.glob(path_alls)))
-        
+
         alls.sort(key=os.path.getmtime)
     else:
         alls = path_alls
-    
+
     pipe = StableDiffusionPipeline.from_pretrained(
         model_id, torch_dtype=torch.float16
     ).to(device)
 
-
     print(f"Found {len(alls)} checkpoints")
     for path in alls[offset:limit]:
         print(path)
@@ -207,8 +204,11 @@ def visualize_progress(
         tune_lora_scale(pipe.text_encoder, text_sclae)
 
         torch.manual_seed(seed)
-        image = pipe(prompt, num_inference_steps=num_inference_steps, guidance_scale=guidance_scale).images[0]
+        image = pipe(
+            prompt,
+            num_inference_steps=num_inference_steps,
+            guidance_scale=guidance_scale,
+        ).images[0]
         imgs.append(image)
 
     return imgs
-

training_scripts/multivector_example.sh

@@ -10,13 +10,16 @@ lora_pti \
   --resolution=512 \
   --train_batch_size=1 \
   --gradient_accumulation_steps=4 \
+  --gradient_checkpointing \
   --scale_lr \
   --learning_rate_unet=1e-4 \
   --learning_rate_text=1e-5 \
   --learning_rate_ti=5e-4 \
   --color_jitter \
   --lr_scheduler="linear" \
   --lr_warmup_steps=0 \
+  --lr_scheduler_lora="linear" \
+  --lr_warmup_steps_lora=100 \
   --placeholder_tokens="<s1>|<s2>" \
   --use_template="style"\
   --save_steps=100 \

training_scripts/train_lora_dreambooth.py

@@ -90,20 +90,25 @@ def __init__(
             self.class_prompt = class_prompt
         else:
             self.class_data_root = None
-
-        img_transforms = []
 
+        img_transforms = []
 
         if resize:
-            img_transforms.append(transforms.Resize(size, interpolation=transforms.InterpolationMode.BILINEAR))
+            img_transforms.append(
+                transforms.Resize(
+                    size, interpolation=transforms.InterpolationMode.BILINEAR
+                )
+            )
         if center_crop:
             img_transforms.append(transforms.CenterCrop(size))
         if color_jitter:
             img_transforms.append(transforms.ColorJitter(0.2, 0.1))
         if h_flip:
             img_transforms.append(transforms.RandomHorizontalFlip())
 
-        self.image_transforms = transforms.Compose([*img_transforms, transforms.ToTensor(), transforms.Normalize([0.5], [0.5])])
+        self.image_transforms = transforms.Compose(
+            [*img_transforms, transforms.ToTensor(), transforms.Normalize([0.5], [0.5])]
+        )
 
     def __len__(self):
         return self._length