app.py

import gradio as gr
import numpy as np
from PIL import Image
from transformers import TimmWrapperModel

import torch
import torchvision.transforms.v2 as T


MODEL_MAP = {
    "p1atdev/style_250412.vit_base_patch16_siglip_384.v2_webli": {
        "mean": [0, 0, 0],
        "std": [1.0, 1.0, 1.0],
        "image_size": 384,
        "background": 0,
    }
}


def config_to_processor(config: dict):
    return T.Compose(
        [
            T.PILToTensor(),
            T.Resize(
                size=None,
                max_size=config["image_size"],
                interpolation=T.InterpolationMode.NEAREST,
            ),
            T.Pad(
                padding=config["image_size"] // 2,
                fill=config["background"],
            ),
            T.CenterCrop(
                size=(config["image_size"], config["image_size"]),
            ),
            T.ToDtype(dtype=torch.float32, scale=True),  # 0~255 -> 0~1
            T.Normalize(mean=config["mean"], std=config["std"]),
        ]
    )


def load_model(name: str):
    return TimmWrapperModel.from_pretrained(name).eval().requires_grad_(False)


MODELS = {
    name: {
        "model": load_model(name),
        "processor": config_to_processor(config),
    }
    for name, config in MODEL_MAP.items()
}


@torch.inference_mode()
def calculate_similarity(model_name: str, image_1: Image.Image, image_2: Image.Image):
    model = MODELS[model_name]["model"]
    processor = MODELS[model_name]["processor"]

    pixel_values = torch.stack([processor(image) for image in [image_1, image_2]])

    embeddings = model(pixel_values).pooler_output
    embeddings /= embeddings.norm(p=2, dim=-1, keepdim=True)

    similarity = (embeddings[0] @ embeddings[1].T).item()

    return similarity


with gr.Blocks() as demo:
    with gr.Row():
        with gr.Column():
            image_1 = gr.Image(label="Image 1", type="pil")
            image_2 = gr.Image(label="Image 2", type="pil")

            model_name = gr.Dropdown(
                label="Model",
                choices=list(MODELS.keys()),
                value=list(MODELS.keys())[0],
            )
            submit_btn = gr.Button("Submit", variant="primary")

        with gr.Column():
            similarity = gr.Label(label="Similarity")

    gr.on(
        triggers=[submit_btn.click],
        fn=calculate_similarity,
        inputs=[
            model_name,
            image_1,
            image_2,
        ],
        outputs=[similarity],
    )

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