app.py

#!/usr/bin/env python

from __future__ import annotations

import argparse
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-Human')

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('--share', action='store_true')
    parser.add_argument('--port', type=int)
    parser.add_argument('--disable-queue',
                        dest='enable_queue',
                        action='store_false')
    return parser.parse_args()


class App:

    def __init__(self, device: torch.device):
        self.device = device
        self.model = self.load_model('stylegan_human_v2_1024.pkl')

    def load_model(self, file_name: str) -> nn.Module:
        path = hf_hub_download('hysts/StyleGAN-Human',
                               f'models/{file_name}',
                               use_auth_token=TOKEN)
        with open(path, 'rb') as f:
            model = pickle.load(f)['G_ema']
        model.eval()
        model.to(self.device)
        with torch.inference_mode():
            z = torch.zeros((1, model.z_dim)).to(self.device)
            label = torch.zeros([1, model.c_dim], device=self.device)
            model(z, label, force_fp32=True)
        return model

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

    @torch.inference_mode()
    def generate_single_image(self, seed: int,
                              truncation_psi: float) -> np.ndarray:
        seed = int(np.clip(seed, 0, np.iinfo(np.uint32).max))

        z = self.generate_z(self.model.z_dim, seed)
        label = torch.zeros([1, self.model.c_dim], device=self.device)

        out = self.model(z,
                         label,
                         truncation_psi=truncation_psi,
                         force_fp32=True)
        out = (out.permute(0, 2, 3, 1) * 127.5 + 128).clamp(0, 255).to(
            torch.uint8)
        return out[0].cpu().numpy()

    @torch.inference_mode()
    def generate_interpolated_images(
            self, seed0: int, psi0: float, seed1: int, psi1: float,
            num_intermediate: int) -> tuple[list[np.ndarray], np.ndarray]:
        seed0 = int(np.clip(seed0, 0, np.iinfo(np.uint32).max))
        seed1 = int(np.clip(seed1, 0, np.iinfo(np.uint32).max))

        z0 = self.generate_z(self.model.z_dim, seed0)
        z1 = self.generate_z(self.model.z_dim, seed1)
        vec = z1 - z0
        dvec = vec / (num_intermediate + 1)
        zs = [z0 + dvec * i for i in range(num_intermediate + 2)]
        dpsi = (psi1 - psi0) / (num_intermediate + 1)
        psis = [psi0 + dpsi * i for i in range(num_intermediate + 2)]

        label = torch.zeros([1, self.model.c_dim], device=self.device)

        res = []
        for z, psi in zip(zs, psis):
            out = self.model(z, label, truncation_psi=psi, force_fp32=True)
            out = (out.permute(0, 2, 3, 1) * 127.5 + 128).clamp(0, 255).to(
                torch.uint8)
            out = out[0].cpu().numpy()
            res.append(out)
        return res


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

    with gr.Blocks(theme=args.theme) as demo:
        gr.Markdown('''<center><h1>StyleGAN-Human</h1></center>

This is a Blocks version of [this app](https://huggingface.co/spaces/hysts/StyleGAN-Human) and [this app](https://huggingface.co/spaces/hysts/StyleGAN-Human-Interpolation).
''')

        with gr.Row():
            with gr.Column():
                with gr.Row():
                    seed1 = gr.Number(value=6876, label='Seed 1')
                    psi1 = gr.Slider(0,
                                     2,
                                     value=0.7,
                                     step=0.05,
                                     label='Truncation psi 1')
                with gr.Row():
                    generate_button1 = gr.Button('Generate')
                with gr.Row():
                    generated_image1 = gr.Image(type='numpy',
                                                label='Generated Image 1')

            with gr.Column():
                with gr.Row():
                    seed2 = gr.Number(value=6886, label='Seed 2')
                    psi2 = gr.Slider(0,
                                     2,
                                     value=0.7,
                                     step=0.05,
                                     label='Truncation psi 2')
                with gr.Row():
                    generate_button2 = gr.Button('Generate')
                with gr.Row():
                    generated_image2 = gr.Image(type='numpy',
                                                label='Generated Image 2')

        with gr.Row():
            with gr.Column():
                with gr.Row():
                    num_frames = gr.Slider(
                        0,
                        41,
                        value=7,
                        step=1,
                        label='Number of Intermediate Frames')
                with gr.Row():
                    interpolate_button = gr.Button('Interpolate')
                with gr.Row():
                    interpolated_images = gr.Gallery(label='Output Images')

        gr.Markdown(
            '<center><img src="https://visitor-badge.glitch.me/badge?page_id=gradio-blocks.stylegan-human" alt="visitor badge"/></center>'
        )

        generate_button1.click(app.generate_single_image,
                               inputs=[seed1, psi1],
                               outputs=generated_image1)
        generate_button2.click(app.generate_single_image,
                               inputs=[seed2, psi2],
                               outputs=generated_image2)
        interpolate_button.click(app.generate_interpolated_images,
                                 inputs=[seed1, psi1, seed2, psi2, num_frames],
                                 outputs=interpolated_images)

    demo.launch(
        enable_queue=args.enable_queue,
        server_port=args.port,
        share=args.share,
    )


if __name__ == '__main__':
    main()