Update for windows compability (#32)

* replaced lambda statements with local functions to work with pickle in dataset.py and moved collate function outside of main and added its own class in cogvideox_image_to_video_lora.py

* Update cogvideox_image_to_video_lora.py

Remove the bug-fix related to image encoding since it's already present in #31

* Update cogvideox_image_to_video_lora.py

Revert to last commit (60c46827404cd4780ee0e7d3c9c56a953e2a04ca)

training/cogvideox_image_to_video_lora.py

@@ -200,6 +200,28 @@ def log_validation(
 
     return videos
 
+class CollateFunction:
+    def __init__(self, weight_dtype, load_tensors):
+        self.weight_dtype = weight_dtype
+        self.load_tensors = load_tensors
+
+    def __call__(self, data):
+        prompts = [x["prompt"] for x in data[0]]
+
+        if self.load_tensors:
+            prompts = torch.stack(prompts).to(dtype=self.weight_dtype, non_blocking=True)
+
+        images = [x["image"] for x in data[0]]
+        images = torch.stack(images).to(dtype=self.weight_dtype, non_blocking=True)
+
+        videos = [x["video"] for x in data[0]]
+        videos = torch.stack(videos).to(dtype=self.weight_dtype, non_blocking=True)
+
+        return {
+            "images": images,
+            "videos": videos,
+            "prompts": prompts,
+        }
 
 def main(args):
     if args.report_to == "wandb" and args.hub_token is not None:
@@ -486,29 +508,13 @@ def main(args):
             video_reshape_mode=args.video_reshape_mode, **dataset_init_kwargs
         )
 
-    def collate_fn(data):
+    collate_fn_instance = CollateFunction(weight_dtype, args.load_tensors)
-        prompts = [x["prompt"] for x in data[0]]
-
-        if args.load_tensors:
-            prompts = torch.stack(prompts).to(dtype=weight_dtype, non_blocking=True)
-
-        images = [x["image"] for x in data[0]]
-        images = torch.stack(images).to(dtype=weight_dtype, non_blocking=True)
-
-        videos = [x["video"] for x in data[0]]
-        videos = torch.stack(videos).to(dtype=weight_dtype, non_blocking=True)
-
-        return {
-            "images": images,
-            "videos": videos,
-            "prompts": prompts,
-        }
 
     train_dataloader = DataLoader(
         train_dataset,
         batch_size=1,
         sampler=BucketSampler(train_dataset, batch_size=args.train_batch_size, shuffle=True),
-        collate_fn=collate_fn,
+        collate_fn=collate_fn_instance,
         num_workers=args.dataloader_num_workers,
         pin_memory=args.pin_memory,
     )
@@ -641,6 +647,7 @@ def main(args):
 
                 # Encode videos
                 if not args.load_tensors:
+                    images = images.permute(0, 2, 1, 3, 4) # [B, C, F, H, W]
                     image_noise_sigma = torch.normal(
                         mean=-3.0, std=0.5, size=(images.size(0),), device=accelerator.device, dtype=weight_dtype
                     )

training/dataset.py

@@ -86,12 +86,20 @@ class VideoDataset(Dataset):
 
         self.video_transforms = transforms.Compose(
             [
-                transforms.RandomHorizontalFlip(random_flip) if random_flip else transforms.Lambda(lambda x: x),
+                transforms.RandomHorizontalFlip(random_flip) if random_flip else transforms.Lambda(self.identity_transform),
-                transforms.Lambda(lambda x: x / 255.0),
+                transforms.Lambda(self.scale_transform),
                 transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True),
             ]
         )
 
+    @staticmethod
+    def identity_transform(x):
+        return x
+
+    @staticmethod
+    def scale_transform(x):
+        return x / 255.0
+
     def __len__(self) -> int:
         return self.num_videos