app.py

#!/usr/bin/env python

import pathlib
import sys

import gradio as gr
import numpy as np
import PIL.Image
import spaces
import torch
import torchvision.transforms as T  # noqa: N812
from huggingface_hub import hf_hub_download

sys.path.insert(0, "CelebAMask-HQ/face_parsing")

from unet import unet  # pyright: ignore[reportMissingImports]
from utils import generate_label  # pyright: ignore[reportMissingImports]

TITLE = "CelebAMask-HQ Face Parsing"
DESCRIPTION = "This is an unofficial demo for the model provided in https://github.com/switchablenorms/CelebAMask-HQ."

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
transform = T.Compose(
    [
        T.Resize((512, 512), interpolation=PIL.Image.NEAREST),
        T.ToTensor(),
        T.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
    ]
)

path = hf_hub_download("public-data/CelebAMask-HQ-Face-Parsing", "models/model.pth")
state_dict = torch.load(path, map_location="cpu")
model = unet()
model.load_state_dict(state_dict)
model.eval()
model.to(device)


@spaces.GPU
@torch.inference_mode()
def predict(image: PIL.Image.Image) -> np.ndarray:
    data = transform(image)
    data = data.unsqueeze(0).to(device)
    out = model(data)
    out = generate_label(out, 512)
    out = out[0].cpu().numpy().transpose(1, 2, 0)
    out = np.clip(np.round(out * 255), 0, 255).astype(np.uint8)

    res = np.asarray(image.resize((512, 512))).astype(float) * 0.5 + out.astype(float) * 0.5
    res = np.clip(np.round(res), 0, 255).astype(np.uint8)
    return out, res


examples = sorted(pathlib.Path("images").glob("*.jpg"))

demo = gr.Interface(
    fn=predict,
    inputs=gr.Image(label="Input", type="pil"),
    outputs=[
        gr.Image(label="Predicted Labels"),
        gr.Image(label="Masked"),
    ],
    examples=examples,
    title=TITLE,
    description=DESCRIPTION,
)

if __name__ == "__main__":
    demo.launch()