Function process_tensor

app.py:176–196

(video_path, fps)

    return video_tensor

def process_tensor(video_path, fps):
    # import decord
    import torchvision.transforms as T
    video_tensor = load_video_with_opencv(video_path)
    # Resize to ensure minimum side is 336
    h, w = video_tensor.shape[2:]
    scale = 336 / min(h, w)
    if scale < 1:
        new_h, new_w = int(h * scale), int(w * scale)
        video_tensor = T.Resize((new_h, new_w))(video_tensor)
    
    video_tensor = video_tensor[::fps].float()[:MAX_FRAMES_OFFLINE]
    
    # Move to GPU
    video_tensor = video_tensor.cuda()
    print(f"Video tensor shape: {video_tensor.shape}, device: {video_tensor.device}")

    # run vggt 
    # process the image tensor
    video_tensor = preprocess_image(video_tensor)[None]
    return video_tensor

def process_and_save_rgb(video_path, user_temp_dir, fps):
    from torchvision.utils import save_image