app.py

import subprocess
from pathlib import Path

import einops
import gradio as gr
import numpy as np
import torch
from huggingface_hub import hf_hub_download
from PIL import Image
from torch import nn
from torchvision.utils import save_image


class Generator(nn.Module):
    def __init__(self, nc=4, nz=100, ngf=64):
        super(Generator, self).__init__()
        self.network = nn.Sequential(
            nn.ConvTranspose2d(nz, ngf * 4, 3, 1, 0, bias=False),
            nn.BatchNorm2d(ngf * 4),
            nn.ReLU(True),
            nn.ConvTranspose2d(ngf * 4, ngf * 2, 3, 2, 1, bias=False),
            nn.BatchNorm2d(ngf * 2),
            nn.ReLU(True),
            nn.ConvTranspose2d(ngf * 2, ngf, 4, 2, 0, bias=False),
            nn.BatchNorm2d(ngf),
            nn.ReLU(True),
            nn.ConvTranspose2d(ngf, nc, 4, 2, 1, bias=False),
            nn.Tanh(),
        )

    def forward(self, input):
        output = self.network(input)
        return output


model = Generator()
weights_path = hf_hub_download('nateraw/cryptopunks-gan', 'generator.pth')
model.load_state_dict(torch.load(weights_path, map_location=torch.device('cpu')))


@torch.no_grad()
def interpolate(save_dir='./lerp/', frames=100, rows=8, cols=8):
    save_dir = Path(save_dir)
    save_dir.mkdir(exist_ok=True, parents=True)

    z1 = torch.randn(rows * cols, 100, 1, 1)
    z2 = torch.randn(rows * cols, 100, 1, 1)

    zs = []
    for i in range(frames):
        alpha = i / frames
        z = (1 - alpha) * z1 + alpha * z2
        zs.append(z)

    zs += zs[::-1]  # also go in reverse order to complete loop

    for i, z in enumerate(zs):
        imgs = model(z)

        # normalize
        imgs = (imgs + 1) / 2

        imgs = (imgs.permute(0, 2, 3, 1).cpu().numpy() * 255).astype(np.uint8)

        # create grid
        imgs = einops.rearrange(imgs, "(b1 b2) h w c -> (b1 h) (b2 w) c", b1=rows, b2=cols)

        Image.fromarray(imgs).save(save_dir / f"{i:03}.png")

    subprocess.call(f"convert -dispose previous -delay 10 -loop 0 {save_dir}/*.png out.gif".split())


def predict(choice, seed):
    torch.manual_seed(seed)

    if choice == 'interpolation':
        interpolate()
        return 'out.gif'
    else:
        z = torch.randn(64, 100, 1, 1)
        punks = model(z)
        save_image(punks, "punks.png", normalize=True)
        return 'punks.png'


gr.Interface(
    predict,
    inputs=[
        gr.inputs.Dropdown(['image', 'interpolation'], label='Output Type'),
        gr.inputs.Slider(label='Seed', minimum=0, maximum=1000, default=42),
    ],
    outputs="image",
    title="Cryptopunks GAN",
    description="These CryptoPunks do not exist. You have the choice of either generating random punks, or a gif showing the interpolation between two random punk grids.",
    article="<p style='text-align: center'><a href='https://arxiv.org/pdf/1511.06434.pdf'>Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks</a> | <a href='https://github.com/teddykoker/cryptopunks-gan'>Github Repo</a></p>",
    examples=[["interpolation", 123], ["interpolation", 42], ["image", 456], ["image", 42]],
).launch(cache_examples=True)