app.py

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

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

def denorm(x):
    return (x * 0.5 + 0.5).clamp(0, 1)

class DCGenerator(nn.Module):
    def __init__(self, z_dim=100, channels=3, f=64):
        super().__init__()
        self.net = nn.Sequential(
            # input is Z
            nn.ConvTranspose2d(z_dim, f * 8, 4, 1, 0, bias=False),
            nn.BatchNorm2d(f * 8),
            nn.ReLU(True),
            #  (f*8) x 4 x 4
            nn.ConvTranspose2d(f * 8, f * 4, 4, 2, 1, bias=False),
            nn.BatchNorm2d(f * 4),
            nn.ReLU(True),
            #  (f*4) x 8 x 8
            nn.ConvTranspose2d(f * 4, f * 2, 4, 2, 1, bias=False),
            nn.BatchNorm2d(f * 2),
            nn.ReLU(True),
            #  (f*2) x 16 x 16
            nn.ConvTranspose2d(f * 2, f, 4, 2, 1, bias=False),
            nn.BatchNorm2d(f),
            nn.ReLU(True),
            #  (f) x 32 x 32
            nn.ConvTranspose2d(f, channels, 4, 2, 1, bias=False),
            nn.Tanh()
            #  (channels) x 64 x 64
        )

    def forward(self, z):
        # print("z shape in generator:", z.shape)
        return self.net(z)

def _strip(sd):
    return {k.replace("module.", ""): v for k, v in sd.items()}

# Download weights from model repo at startup
dc_path = hf_hub_download(repo_id="SotaSF/q1-models", filename="dcgan_final.pt")
wg_path = hf_hub_download(repo_id="SotaSF/q1-models", filename="wgangp_final.pt")

@torch.no_grad()
def generate_compare(n_samples, seed, z_dim):
    g1 = DCGenerator(int(z_dim)).to(device)
    g2 = DCGenerator(int(z_dim)).to(device)

    s1 = torch.load(dc_path, map_location=device)
    s2 = torch.load(wg_path, map_location=device)

    g1.load_state_dict(_strip(s1["G"] if "G" in s1 else s1))
    g2.load_state_dict(_strip(s2["G"] if "G" in s2 else s2))

    g1.eval()
    g2.eval()

    gen = torch.Generator(device=device).manual_seed(int(seed))
    z = torch.randn(int(n_samples), int(z_dim), 1, 1, device=device, generator=gen)

    a = denorm(g1(z)).cpu()
    b = denorm(g2(z)).cpu()

    out_a, out_b = [], []
    for i in range(int(n_samples)):
        out_a.append((a[i].permute(1,2,0).numpy()*255).astype(np.uint8))
        out_b.append((b[i].permute(1,2,0).numpy()*255).astype(np.uint8))

    return out_a, out_b

with gr.Blocks(title="Q1 DCGAN vs WGAN-GP") as demo:
    gr.Markdown("# Q1: DCGAN vs WGAN-GP")
    n = gr.Slider(5, 20, value=10, step=1, label="Samples")
    s = gr.Number(value=42, label="Seed", precision=0)
    z = gr.Number(value=100, label="z dim", precision=0)
    btn = gr.Button("Generate")

    g1 = gr.Gallery(label="DCGAN", columns=5)
    g2 = gr.Gallery(label="WGAN-GP", columns=5)

    btn.click(generate_compare, [n, s, z], [g1, g2])

demo.launch()