app.py

#!/usr/bin/env python

from __future__ import annotations

import argparse
import functools
import os
import pickle
import sys

import gradio as gr
import numpy as np
import torch
import torch.nn as nn
from huggingface_hub import hf_hub_download

sys.path.insert(0, 'stylegan_xl')

ORIGINAL_REPO_URL = 'https://github.com/autonomousvision/stylegan_xl'
TITLE = 'autonomousvision/stylegan_xl'
DESCRIPTION = f'''This is a demo for {ORIGINAL_REPO_URL}.

For class-conditional models, you can specify the class index.
Index-to-label dictionaries for ImageNet and CIFAR-10 can be found [here](https://raw.githubusercontent.com/autonomousvision/stylegan_xl/main/misc/imagenet_idx2labels.txt) and [here](https://www.cs.toronto.edu/~kriz/cifar.html), respectively.
'''
SAMPLE_IMAGE_DIR = 'https://huggingface.co/spaces/hysts/StyleGAN-XL/resolve/main/samples'
ARTICLE = f'''## Generated images
- truncation: 0.7
### ImageNet
- size: 128x128
- class index: 0-999
- seed: 0
![ImageNet samples]({SAMPLE_IMAGE_DIR}/imagenet.jpg)
### CIFAR-10
- size: 32x32
- class index: 0-9
- seed: 0-9
![CIFAR-10 samples]({SAMPLE_IMAGE_DIR}/cifar10.jpg)
### FFHQ
- size: 256x256
- seed: 0-99
![FFHQ samples]({SAMPLE_IMAGE_DIR}/ffhq.jpg)
### Pokemon
- size: 256x256
- seed: 0-99
![Pokemon samples]({SAMPLE_IMAGE_DIR}/pokemon.jpg)
'''

TOKEN = os.environ['TOKEN']


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser()
    parser.add_argument('--device', type=str, default='cpu')
    parser.add_argument('--theme', type=str)
    parser.add_argument('--live', action='store_true')
    parser.add_argument('--share', action='store_true')
    parser.add_argument('--port', type=int)
    parser.add_argument('--disable-queue',
                        dest='enable_queue',
                        action='store_false')
    parser.add_argument('--allow-flagging', type=str, default='never')
    return parser.parse_args()


def make_transform(translate: tuple[float, float], angle: float) -> np.ndarray:
    mat = np.eye(3)
    sin = np.sin(angle / 360 * np.pi * 2)
    cos = np.cos(angle / 360 * np.pi * 2)
    mat[0][0] = cos
    mat[0][1] = sin
    mat[0][2] = translate[0]
    mat[1][0] = -sin
    mat[1][1] = cos
    mat[1][2] = translate[1]
    return mat


def generate_z(z_dim: int, seed: int, device: torch.device) -> torch.Tensor:
    return torch.from_numpy(np.random.RandomState(seed).randn(
        1, z_dim)).to(device).float()


@torch.inference_mode()
def generate_image(model_name: str, class_index: int, seed: int,
                   truncation_psi: float, tx: float, ty: float, angle: float,
                   model_dict: dict[str, nn.Module],
                   device: torch.device) -> np.ndarray:
    model = model_dict[model_name]
    seed = int(np.clip(seed, 0, np.iinfo(np.uint32).max))

    z = generate_z(model.z_dim, seed, device)

    label = torch.zeros([1, model.c_dim], device=device)
    class_index = round(class_index)
    class_index = min(max(0, class_index), model.c_dim - 1)
    class_index = torch.tensor(class_index, dtype=torch.long)
    if class_index >= 0:
        label[:, class_index] = 1

    mat = make_transform((tx, ty), angle)
    mat = np.linalg.inv(mat)
    model.synthesis.input.transform.copy_(torch.from_numpy(mat))

    out = model(z, label, truncation_psi=truncation_psi)
    out = (out.permute(0, 2, 3, 1) * 127.5 + 128).clamp(0, 255).to(torch.uint8)
    return out[0].cpu().numpy()


def load_model(model_name: str, device: torch.device) -> nn.Module:
    path = hf_hub_download('hysts/StyleGAN-XL',
                           f'models/{model_name}.pkl',
                           use_auth_token=TOKEN)
    with open(path, 'rb') as f:
        model = pickle.load(f)['G_ema']
    model.eval()
    model.to(device)
    with torch.inference_mode():
        z = torch.zeros((1, model.z_dim)).to(device)
        label = torch.zeros([1, model.c_dim], device=device)
        model(z, label)
    return model


def main():
    args = parse_args()
    device = torch.device(args.device)

    model_names = [
        'imagenet16',
        'imagenet32',
        'imagenet64',
        'imagenet128',
        'cifar10',
        'ffhq256',
        'pokemon256',
    ]

    model_dict = {name: load_model(name, device) for name in model_names}

    func = functools.partial(generate_image,
                             model_dict=model_dict,
                             device=device)
    func = functools.update_wrapper(func, generate_image)

    gr.Interface(
        func,
        [
            gr.inputs.Radio(model_names,
                            type='value',
                            default='imagenet128',
                            label='Model'),
            gr.inputs.Number(default=284, label='Class index'),
            gr.inputs.Number(default=0, label='Seed'),
            gr.inputs.Slider(
                0, 2, step=0.05, default=0.7, label='Truncation psi'),
            gr.inputs.Slider(-1, 1, step=0.05, default=0, label='Translate X'),
            gr.inputs.Slider(-1, 1, step=0.05, default=0, label='Translate Y'),
            gr.inputs.Slider(-180, 180, step=5, default=0, label='Angle'),
        ],
        gr.outputs.Image(type='numpy', label='Output'),
        title=TITLE,
        description=DESCRIPTION,
        article=ARTICLE,
        theme=args.theme,
        allow_flagging=args.allow_flagging,
        live=args.live,
    ).launch(
        enable_queue=args.enable_queue,
        server_port=args.port,
        share=args.share,
    )


if __name__ == '__main__':
    main()