app.py

#!/usr/bin/env python

import pathlib
import sys

import cv2
import gradio as gr
import numpy as np
import spaces
import torch
from huggingface_hub import hf_hub_download
from torch import nn

submodule_dir = pathlib.Path(__file__).parent / "MangaLineExtraction_PyTorch"
sys.path.append(submodule_dir.as_posix())

from model_torch import res_skip  # type: ignore # noqa: E402

DESCRIPTION = "# [MangaLineExtraction_PyTorch](https://github.com/ljsabc/MangaLineExtraction_PyTorch)"


def load_model(device: torch.device) -> nn.Module:
    ckpt_path = hf_hub_download("public-data/MangaLineExtraction_PyTorch", "erika.pth")
    state_dict = torch.load(ckpt_path)
    model = res_skip()
    model.load_state_dict(state_dict)
    model.to(device)
    model.eval()
    return model


MAX_SIZE = 1000
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = load_model(device)


@spaces.GPU
@torch.inference_mode()
def predict(image: np.ndarray) -> np.ndarray:
    gray = cv2.cvtColor(image, cv2.COLOR_RGB2GRAY)

    if max(gray.shape) > MAX_SIZE:
        scale = MAX_SIZE / max(gray.shape)
        gray = cv2.resize(gray, None, fx=scale, fy=scale)

    h, w = gray.shape
    size = 16
    new_w = (w + size - 1) // size * size
    new_h = (h + size - 1) // size * size

    patch = np.ones((1, 1, new_h, new_w), dtype=np.float32)
    patch[0, 0, :h, :w] = gray
    tensor = torch.from_numpy(patch).to(device)
    out = model(tensor)

    res = out.cpu().numpy()[0, 0, :h, :w]
    return np.clip(res, 0, 255).astype(np.uint8)


with gr.Blocks(css_paths="style.css") as demo:
    gr.Markdown(DESCRIPTION)
    with gr.Row():
        with gr.Column():
            input_image = gr.Image(label="Input", type="numpy")
            run_button = gr.Button()
        with gr.Column():
            result = gr.Image(label="Result", elem_id="result")
    run_button.click(
        fn=predict,
        inputs=input_image,
        outputs=result,
    )

if __name__ == "__main__":
    demo.queue(max_size=20).launch()