app.py

#!/usr/bin/env python

import pathlib
import sys
import tarfile

import cv2
import gradio as gr
import huggingface_hub
import numpy as np
import PIL.Image
import spaces
import torch

sys.path.insert(0, "yolov5_anime")

from models.yolo import Model  # pyright: ignore[reportMissingImports]
from utils.datasets import letterbox  # pyright: ignore[reportMissingImports]
from utils.general import non_max_suppression, scale_coords  # pyright: ignore[reportMissingImports]

DESCRIPTION = "# [zymk9/yolov5_anime](https://github.com/zymk9/yolov5_anime)"

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

torch.set_grad_enabled(False)
MODEL_REPO = "public-data/yolov5_anime"
model_path = huggingface_hub.hf_hub_download(MODEL_REPO, "yolov5x_anime.pth")
config_path = huggingface_hub.hf_hub_download(MODEL_REPO, "yolov5x.yaml")
state_dict = torch.load(model_path, weights_only=True)
model = Model(cfg=config_path)
model.load_state_dict(state_dict)
if device.type != "cpu":
    model.half()
model.to(device)
model.eval()


def load_sample_image_paths() -> list[pathlib.Path]:
    image_dir = pathlib.Path("images")
    if not image_dir.exists():
        dataset_repo = "hysts/sample-images-TADNE"
        path = huggingface_hub.hf_hub_download(dataset_repo, "images.tar.gz", repo_type="dataset")
        with tarfile.open(path) as f:
            f.extractall()  # noqa: S202
    return sorted(image_dir.glob("*"))


@spaces.GPU
@torch.inference_mode()
def predict(image: PIL.Image.Image, score_threshold: float, iou_threshold: float) -> np.ndarray:
    orig_image = np.asarray(image)

    image = letterbox(orig_image, new_shape=640)[0]
    data = torch.from_numpy(image.transpose(2, 0, 1)).float() / 255
    data = data.to(device).unsqueeze(0)
    if device.type != "cpu":
        data = data.half()

    preds = model(data)[0]
    preds = non_max_suppression(preds, score_threshold, iou_threshold)

    detections = []
    for pred in preds:
        if pred is not None and len(pred) > 0:
            pred[:, :4] = scale_coords(data.shape[2:], pred[:, :4], orig_image.shape).round()
            # (x0, y0, x1, y0, conf, class)
            detections.append(pred.cpu().numpy())
    detections = np.concatenate(detections) if detections else np.empty(shape=(0, 6))

    res = orig_image.copy()
    for det in detections:
        x0, y0, x1, y1 = det[:4].astype(int)
        cv2.rectangle(res, (x0, y0), (x1, y1), (0, 255, 0), 3)
    return res


image_paths = load_sample_image_paths()
examples = [[path.as_posix(), 0.4, 0.5] for path in image_paths]


with gr.Blocks() as demo:
    gr.Markdown(DESCRIPTION)
    with gr.Row():
        with gr.Column():
            image = gr.Image(label="Input", type="pil")
            score_threshold = gr.Slider(label="Score Threshold", minimum=0, maximum=1, step=0.05, value=0.4)
            iou_threshold = gr.Slider(label="IoU Threshold", minimum=0, maximum=1, step=0.05, value=0.5)
            run_button = gr.Button()
        with gr.Column():
            result = gr.Image(label="Output")

    inputs = [image, score_threshold, iou_threshold]
    gr.Examples(
        examples=examples,
        inputs=inputs,
        outputs=result,
        fn=predict,
    )
    run_button.click(
        fn=predict,
        inputs=inputs,
        outputs=result,
    )

if __name__ == "__main__":
    demo.launch(css_paths="style.css")