Spaces:
Build error
Build error
| import gc | |
| import gradio as gr | |
| import numpy as np | |
| import torch | |
| from huggingface_hub import hf_hub_download | |
| from PIL.Image import Resampling | |
| from pytorchvideo.data.encoded_video import EncodedVideo | |
| from pytorchvideo.transforms.functional import uniform_temporal_subsample | |
| from torchvision.io import write_video | |
| from torchvision.transforms.functional import resize | |
| from modeling import Generator | |
| MAX_DURATION = 4 | |
| OUT_FPS = 18 | |
| DEVICE = "cpu" if not torch.cuda.is_available() else "cuda" | |
| # Reupload of model found here: https://huggingface.co/spaces/awacke1/Image2LineDrawing | |
| model = Generator(3, 1, 3) | |
| weights_path = hf_hub_download("nateraw/image-2-line-drawing", "pytorch_model.bin") | |
| model.load_state_dict(torch.load(weights_path, map_location=DEVICE)) | |
| model.eval() | |
| def process_one_second(vid, start_sec, out_fps): | |
| """Process one second of a video at a given fps | |
| Args: | |
| vid (_type_): A pytorchvideo.EncodedVideo instance containing the video to process | |
| start_sec (_type_): The second to start processing at | |
| out_fps (_type_): The fps to output the video at | |
| Returns: | |
| np.array: The processed video as a numpy array with shape (T, H, W, C) | |
| """ | |
| # C, T, H, W | |
| video_arr = vid.get_clip(start_sec, start_sec + 1)["video"] | |
| # C, T, H, W where T == frames per second | |
| x = uniform_temporal_subsample(video_arr, out_fps) | |
| # C, T, H, W where H has been scaled to 256 (This will probably be no bueno on vertical vids but whatever) | |
| x = resize(x, 256, Resampling.BICUBIC) | |
| # C, T, H, W -> T, C, H, W (basically T acts as batch size now) | |
| x = x.permute(1, 0, 2, 3) | |
| with torch.no_grad(): | |
| # T, 1, H, W | |
| out = model(x) | |
| # T, C, H, W -> T, H, W, C Rescaled to 0-255 | |
| out = out.permute(0, 2, 3, 1).clip(0, 1) * 255 | |
| # Greyscale -> RGB | |
| out = out.repeat(1, 1, 1, 3) | |
| return out | |
| def fn(fpath): | |
| start_sec = 0 | |
| vid = EncodedVideo.from_path(fpath) | |
| duration = min(MAX_DURATION, int(vid.duration)) | |
| for i in range(duration): | |
| print(f"🖼️ Processing step {i + 1}/{duration}...") | |
| video = process_one_second(vid, start_sec=i + start_sec, out_fps=OUT_FPS) | |
| gc.collect() | |
| if i == 0: | |
| video_all = video | |
| else: | |
| video_all = np.concatenate((video_all, video)) | |
| write_video("out.mp4", video_all, fps=OUT_FPS) | |
| return "out.mp4" | |
| webcam_interface = gr.Interface( | |
| fn, gr.Video(source="webcam"), gr.Video(type="file", format="mp4") | |
| ) | |
| video_interface = gr.Interface( | |
| fn, gr.Video(type="file"), gr.Video(type="file", format="mp4") | |
| ) | |
| if __name__ == '__main__': | |
| gr.TabbedInterface( | |
| [webcam_interface, video_interface], | |
| ["Run on Your Webcam!", "Run on Videos!"], | |
| ).launch() | |