app.py

#!/usr/bin/env python

from __future__ import annotations

import io
import pathlib
import tarfile

import deepdanbooru as dd
import gradio as gr
import huggingface_hub
import numpy as np
import PIL.Image
import tensorflow as tf
from huggingface_hub import hf_hub_download

TITLE = "TADNE Image Search with DeepDanbooru"
DESCRIPTION = """The original TADNE site is https://thisanimedoesnotexist.ai/.

This app shows images similar to the query image from images generated
by the TADNE model with seed 0-99999.
Here, image similarity is measured by the L2 distance of the intermediate
features by the [DeepDanbooru](https://github.com/KichangKim/DeepDanbooru)
model.

The resolution of the output images in this app is 128x128, but you can
check the original 512x512 images from URLs like
https://thisanimedoesnotexist.ai/slider.html?seed=10000 using the output seeds.

Expected execution time on Hugging Face Spaces: 7s

Related Apps:
- [TADNE](https://huggingface.co/spaces/hysts/TADNE)
- [TADNE Image Viewer](https://huggingface.co/spaces/hysts/TADNE-image-viewer)
- [TADNE Image Selector](https://huggingface.co/spaces/hysts/TADNE-image-selector)
- [TADNE Interpolation](https://huggingface.co/spaces/hysts/TADNE-interpolation)
- [DeepDanbooru](https://huggingface.co/spaces/hysts/DeepDanbooru)
"""


def load_deepdanbooru_predictions(dirname: str) -> np.ndarray:
    path = hf_hub_download(
        "hysts/TADNE-sample-images",
        f"prediction_results/deepdanbooru/intermediate_features/{dirname}.npy",
        repo_type="dataset",
    )
    return np.load(path)


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()
    return sorted(image_dir.glob("*"))


def create_model() -> tf.keras.Model:
    path = huggingface_hub.hf_hub_download("public-data/DeepDanbooru", "model-resnet_custom_v3.h5")
    model = tf.keras.models.load_model(path)
    model = tf.keras.Model(model.input, model.layers[-4].output)
    layer = tf.keras.layers.GlobalAveragePooling2D()
    model = tf.keras.Sequential([model, layer])
    return model


image_size = 128
dirname = "0-99999"
tarball_path = hf_hub_download("hysts/TADNE-sample-images", f"{image_size}/{dirname}.tar", repo_type="dataset")
deepdanbooru_predictions = load_deepdanbooru_predictions(dirname)

model = create_model()


def predict(image: PIL.Image.Image) -> np.ndarray:
    _, height, width, _ = model.input_shape
    image = np.asarray(image)
    image = tf.image.resize(image, size=(height, width), method=tf.image.ResizeMethod.AREA, preserve_aspect_ratio=True)
    image = image.numpy()
    image = dd.image.transform_and_pad_image(image, width, height)
    image = image / 255.0
    features = model.predict(image[None, ...])[0]
    features = features.astype(float)
    return features


def run(
    image: PIL.Image.Image,
    nrows: int,
    ncols: int,
) -> tuple[np.ndarray, np.ndarray]:
    features = predict(image)
    distances = ((deepdanbooru_predictions - features) ** 2).sum(axis=1)

    image_indices = np.argsort(distances)

    seeds = []
    images = []
    with tarfile.TarFile(tarball_path) as tar_file:
        for index in range(nrows * ncols):
            image_index = image_indices[index]
            seeds.append(image_index)
            member = tar_file.getmember(f"{dirname}/{image_index:07d}.jpg")
            with tar_file.extractfile(member) as f:  # type: ignore
                data = io.BytesIO(f.read())
            image = PIL.Image.open(data)
            image = np.asarray(image)
            images.append(image)
    res = (
        np.asarray(images)
        .reshape(nrows, ncols, image_size, image_size, 3)
        .transpose(0, 2, 1, 3, 4)
        .reshape(nrows * image_size, ncols * image_size, 3)
    )
    seeds = np.asarray(seeds).reshape(nrows, ncols)

    return res, seeds


image_paths = load_sample_image_paths()
examples = [[path.as_posix(), 2, 5] for path in image_paths]

demo = gr.Interface(
    fn=run,
    inputs=[
        gr.Image(label="Input", type="pil"),
        gr.Slider(label="Number of Rows", minimum=1, maximum=10, step=1, value=2),
        gr.Slider(label="Number of Columns", minimum=1, maximum=10, step=1, value=2),
    ],
    outputs=[
        gr.Image(label="Output"),
        gr.Dataframe(label="Seed"),
    ],
    examples=examples,
    title=TITLE,
    description=DESCRIPTION,
)


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