app.py

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()