app.py

#!/usr/bin/env python

import functools
import os
import pathlib
import sys
import tarfile
from collections.abc import Callable

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

sys.path.insert(0, "bizarre-pose-estimator")

from _util.twodee_v0 import I as ImageWrapper

DESCRIPTION = (
    "# [ShuhongChen/bizarre-pose-estimator (segmenter)](https://github.com/ShuhongChen/bizarre-pose-estimator)"
)


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("*"))


def load_model(device: torch.device) -> tuple[torch.nn.Module, torch.nn.Module]:
    path = huggingface_hub.hf_hub_download("public-data/bizarre-pose-estimator-models", "segmenter.pth")
    ckpt = torch.load(path)

    model = torchvision.models.segmentation.deeplabv3_resnet101()
    model.classifier = nn.Sequential(
        torchvision.models.segmentation.deeplabv3.ASPP(2048, [12, 24, 36]),
        nn.Conv2d(256, 64, kernel_size=3, stride=1, padding=1),
        nn.BatchNorm2d(64),
        nn.LeakyReLU(),
        nn.Conv2d(64, 16, kernel_size=3, stride=1, padding=1),
        nn.BatchNorm2d(16),
        nn.LeakyReLU(),
    )
    final_head = nn.Sequential(
        nn.Conv2d(16 + 3, 16, kernel_size=3, stride=1, padding=1),
        nn.BatchNorm2d(16),
        nn.LeakyReLU(),
        nn.Conv2d(16, 8, kernel_size=3, stride=1, padding=1),
        nn.BatchNorm2d(8),
        nn.LeakyReLU(),
        nn.Conv2d(8, 2, kernel_size=1, stride=1),
    )
    model.load_state_dict(ckpt["model"])
    final_head.load_state_dict(ckpt["final_head"])
    model.to(device)
    model.eval()
    final_head.to(device)
    final_head.eval()
    return model, final_head


@torch.inference_mode()
def predict(
    image: PIL.Image.Image,
    score_threshold: float,
    transform: Callable,
    device: torch.device,
    model: torch.nn.Module,
    final_head: torch.nn.Module,
) -> np.ndarray:
    data = ImageWrapper(image).resize_min(256).convert("RGBA").alpha_bg(1).convert("RGB").pil()
    data = torchvision.transforms.functional.to_tensor(data)
    data = transform(data)
    data = data.to(device).unsqueeze(0)

    out = model(data)["out"]
    out_fin = final_head(
        torch.cat(
            [
                out,
                data,
            ],
            dim=1,
        )
    )
    probs = torch.softmax(out_fin, dim=1)[0]
    probs = probs[1]  # foreground
    probs = PIL.Image.fromarray(probs.cpu().numpy()).resize(image.size)

    mask = np.asarray(probs).copy()
    mask[mask < score_threshold] = 0
    mask[mask > 0] = 1
    mask = mask.astype(bool)

    res = np.asarray(image).copy()
    res[~mask] = 255
    return res


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

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model, final_head = load_model(device)
transform = T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])

fn = functools.partial(predict, transform=transform, device=device, model=model, final_head=final_head)

with gr.Blocks(css_paths="style.css") as demo:
    gr.Markdown(DESCRIPTION)
    with gr.Row():
        with gr.Column():
            image = gr.Image(label="Input", type="pil")
            threshold = gr.Slider(label="Score Threshold", minimum=0, maximum=1, step=0.05, value=0.5)
            run_button = gr.Button("Run")
        with gr.Column():
            result = gr.Image(label="Masked")

    inputs = [image, threshold]
    gr.Examples(
        examples=examples,
        inputs=inputs,
        outputs=result,
        fn=fn,
        cache_examples=os.getenv("CACHE_EXAMPLES") == "1",
    )
    run_button.click(
        fn=fn,
        inputs=inputs,
        outputs=result,
        api_name="predict",
    )

if __name__ == "__main__":
    demo.queue(max_size=15).launch()