Open-Sora: Democratizing Efficient Video Production for All
Python
fix dataset loading
Shen Chenhuib5a3334cafa9
diff --git a/configs/vae_magvit_v2/train/16x128x128_pixabay_8_GPU_pipeline.py b/configs/vae_magvit_v2/train/16x128x128_pixabay_8_GPU_pipeline.pyindex 31af832..7b2c60c 100644--- a/configs/vae_magvit_v2/train/16x128x128_pixabay_8_GPU_pipeline.py+++ b/configs/vae_magvit_v2/train/16x128x128_pixabay_8_GPU_pipeline.py@@ -4,6 +4,7 @@ dataset = dict(num_frames=16,frame_interval=3,image_size=(128, 128),+ get_text=False,)# Define accelerationdiff --git a/opensora/datasets/datasets.py b/opensora/datasets/datasets.pyindex 4c2b316..d320ff7 100644--- a/opensora/datasets/datasets.py+++ b/opensora/datasets/datasets.py@@ -29,6 +29,7 @@ class VideoTextDataset(torch.utils.data.Dataset):frame_interval=1,image_size=(256, 256),transform_name="center",+ get_text=True,):self.data_path = data_pathself.data = read_file(data_path)@@ -39,6 +40,7 @@ class VideoTextDataset(torch.utils.data.Dataset):"image": get_transforms_image(transform_name, image_size),"video": get_transforms_video(transform_name, image_size),}+ self.get_text = get_textdef _print_data_number(self):num_videos = 0@@ -61,7 +63,7 @@ class VideoTextDataset(torch.utils.data.Dataset):def getitem(self, index):sample = self.data.iloc[index]path = sample["path"]- text = sample["text"]+ text = sample["text"] if self.get_text else Nonefile_type = self.get_type(path)if file_type == "video":@@ -87,7 +89,11 @@ class VideoTextDataset(torch.utils.data.Dataset):# TCHW -> CTHWvideo = video.permute(1, 0, 2, 3)- return {"video": video, "text": text}++ if self.get_text:+ return {"video": video, "text": text}+ else:+ return {"video": video}def __getitem__(self, index):for _ in range(10):