app.py

import os
import numpy as np
from PIL import Image
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")
CKPT_PATH = hf_hub_download(repo_id="SotaSF/q3-model", filename="cyclegan_final.pt")

class ResnetBlock(nn.Module):
    def __init__(self, c):
        super().__init__()
        self.b = nn.Sequential(
            nn.ReflectionPad2d(1),
            nn.Conv2d(c, c, 3),
            nn.InstanceNorm2d(c),
            nn.ReLU(True),
            nn.ReflectionPad2d(1),
            nn.Conv2d(c, c, 3),
            nn.InstanceNorm2d(c),
        )

    def forward(self, x):
        return x + self.b(x)

class ResnetGenerator(nn.Module):
    def __init__(self, in_c=3, out_c=3, n_blocks=6, base=64):
        super().__init__()
        # c7s1-64
        m = [nn.ReflectionPad2d(3), nn.Conv2d(in_c, base, 7), nn.InstanceNorm2d(base), nn.ReLU(True)]
        f = base
        # d128 and d256
        for _ in range(2):
            m += [nn.Conv2d(f, f * 2, 3, 2, 1), nn.InstanceNorm2d(f * 2), nn.ReLU(True)]
            f *= 2
        # R256
        for _ in range(n_blocks):
            m += [ResnetBlock(f)]
        # u128 and u64
        for _ in range(2):
            m += [nn.ConvTranspose2d(f, f // 2, 3, 2, 1, output_padding=1), nn.InstanceNorm2d(f // 2), nn.ReLU(True)]
            f //= 2
        # c7s1-3
        m += [nn.ReflectionPad2d(3), nn.Conv2d(base, out_c, 7), nn.Tanh()]
        self.m = nn.Sequential(*m)

    def forward(self, x):
        return self.m(x)

def clean_sd(sd):
    out = {}
    for k, v in sd.items():
        if k.startswith("module."):
            out[k[7:]] = v
        else:
            out[k] = v
    return out

g_ab = None
g_ba = None
load_error = None

def load_models():
    global g_ab, g_ba, load_error
    if g_ab is not None or load_error is not None:
        return
    try:
        sd = torch.load(CKPT_PATH, map_location=device)
        g_ab = ResnetGenerator(n_blocks=6).to(device)
        g_ba = ResnetGenerator(n_blocks=6).to(device)
        g_ab.load_state_dict(clean_sd(sd["G_AB"]), strict=False)
        g_ba.load_state_dict(clean_sd(sd["G_BA"]), strict=False)
        g_ab.eval()
        g_ba.eval()
    except Exception as e:
        load_error = str(e)

def preprocess(inp, size=128):
    img = Image.fromarray(inp.astype(np.uint8)).convert("RGB").resize((size, size))
    arr = np.asarray(img).astype(np.float32) / 255.0
    arr = (arr - 0.5) / 0.5
    arr = np.transpose(arr, (2, 0, 1))
    x = torch.from_numpy(arr).unsqueeze(0)
    return x

@torch.no_grad()
def run(inp, direction='Sketch -> Photo (G_AB)'):
    load_models()
    if inp is None:
        return None
    if load_error is not None:
        raise RuntimeError("Model load failed: " + load_error)
    x = preprocess(inp, 128).to(device)
    y = g_ab(x) if direction.startswith("Sketch") else g_ba(x)
    y = (y * 0.5 + 0.5).clamp(0, 1)
    return (y[0].permute(1, 2, 0).cpu().numpy() * 255).astype(np.uint8)

with gr.Blocks(title="Q3: CycleGAN") as demo:
    gr.Markdown("# Q3: CycleGAN")
    inp = gr.Image(type="numpy", label="Input")
    direction = gr.Radio(["Sketch -> Photo (G_AB)", "Photo -> Sketch (G_BA)"], value="Sketch -> Photo (G_AB)", label="Direction")
    out = gr.Image(type="numpy", label="Output")
    btn = gr.Button("Translate")
    btn.click(run, [inp, direction], [out])

demo.launch(server_name="0.0.0.0", server_port=7860)