app.py

import os
import gradio as gr
import numpy as np
import torch
import torch.nn.functional as F
from torchvision import transforms
from PIL import Image
import tempfile

from gradio.themes.utils import sizes
from classes_and_palettes import GOLIATH_CLASSES


# =========================================================
# Config
# =========================================================

class Config:
    ASSETS_DIR = os.path.join(os.path.dirname(__file__), "assets")
    CHECKPOINTS_DIR = os.path.join(ASSETS_DIR, "checkpoints")
    CHECKPOINTS = {
        "0.3b": "sapiens_0.3b_goliath_best_goliath_mIoU_7673_epoch_194_torchscript.pt2",
        "0.6b": "sapiens_0.6b_goliath_best_goliath_mIoU_7777_epoch_178_torchscript.pt2",
        "1b": "sapiens_1b_goliath_best_goliath_mIoU_7994_epoch_151_torchscript.pt2",
    }


# =========================================================
# Model
# =========================================================

class ModelManager:
    _cache = {}

    @staticmethod
    def load_model(name: str):
        if name in ModelManager._cache:
            return ModelManager._cache[name]

        path = os.path.join(Config.CHECKPOINTS_DIR, Config.CHECKPOINTS[name])
        model = torch.jit.load(path)
        model.eval().to("cuda")
        ModelManager._cache[name] = model
        return model

    @staticmethod
    @torch.inference_mode()
    def run(model, x, h, w):
        out = model(x)
        out = F.interpolate(out, size=(h, w), mode="bilinear", align_corners=False)
        return out.argmax(1)


# =========================================================
# Image Processing
# =========================================================

class ImageProcessor:
    def __init__(self):
        self.tf = transforms.Compose([
            transforms.Resize((1024, 768)),
            transforms.ToTensor(),
            transforms.Normalize(
                mean=[123.5 / 255, 116.5 / 255, 103.5 / 255],
                std=[58.5 / 255, 57.0 / 255, 57.5 / 255],
            ),
        ])

    def process(self, image: Image.Image, model_name: str):
        model = ModelManager.load_model(model_name)
        x = self.tf(image).unsqueeze(0).to("cuda")

        pred = ModelManager.run(model, x, image.height, image.width)
        mask = pred.squeeze(0).cpu().numpy()

        # Save raw mask
        npy_path = tempfile.mktemp(suffix=".npy")
        np.save(npy_path, mask)

        # Build AnnotatedImage output
        annotations = self._build_annotations(mask)

        return (image, annotations), npy_path

    def _build_annotations(self, mask: np.ndarray):
        annotations = []
        for class_id in np.unique(mask):
            if class_id >= len(GOLIATH_CLASSES):
                continue

            binary_mask = (mask == class_id).astype(np.uint8)
            if binary_mask.sum() == 0:
                continue

            annotations.append(
                (binary_mask, GOLIATH_CLASSES[class_id])
            )

        return annotations


# =========================================================
# UI
# =========================================================

class GradioInterface:
    def __init__(self):
        self.processor = ImageProcessor()

    def create(self):
        def run(image, model):
            return self.processor.process(image, model)

        with gr.Blocks() as demo:
            with gr.Row():
                with gr.Column(scale=1):
                    input_image = gr.Image(
                        label="Input Image",
                        type="pil",
                    )

                    model_name = gr.Dropdown(
                        label="Model Size",
                        choices=list(Config.CHECKPOINTS.keys()),
                        value="1b",
                    )

                    run_btn = gr.Button("Run Segmentation", variant="primary")

                with gr.Column(scale=2):
                    annotated = gr.AnnotatedImage(
                        label="Segmentation Result",
                        show_legend=True,
                        height=512,
                    )

                    mask_file = gr.File(label="Raw Mask (.npy)")

            run_btn.click(
                fn=run,
                inputs=[input_image, model_name],
                outputs=[annotated, mask_file],
            )

        return demo


# =========================================================
# Entrypoint
# =========================================================

def main():
    if torch.cuda.is_available() and torch.cuda.get_device_properties(0).major >= 8:
        torch.backends.cuda.matmul.allow_tf32 = True
        torch.backends.cudnn.allow_tf32 = True

    app = GradioInterface().create()
    app.launch(server_name="0.0.0.0", share=False)


if __name__ == "__main__":
    main()