source/app.py

from __future__ import annotations

from pathlib import Path

import gradio as gr
import torch
from model import ConditionalEnergyNetwork, langevin_sample
from PIL import Image
from safetensors.torch import load_file

ARTIFACT_DIR = Path(__file__).resolve().parent / "artifacts" / "energy-pocket"
MODEL = ConditionalEnergyNetwork()
MODEL.load_state_dict(load_file(ARTIFACT_DIR / "model.safetensors"))
MODEL.eval()


def sample_energy(
    label: int,
    seed: int,
    steps: int,
) -> tuple[Image.Image, dict]:
    generator = torch.Generator().manual_seed(int(seed))
    labels = torch.full((12,), int(label), dtype=torch.long)
    initial = torch.rand(12, 64, generator=generator)
    generated = langevin_sample(
        MODEL,
        initial,
        labels,
        steps=int(steps),
        step_size=0.08,
        noise_scale=0.008,
        generator=generator,
    )
    canvas = Image.new("L", (512, 384), color=0)
    for index, pixels in enumerate(generated):
        image = Image.fromarray(
            pixels.reshape(8, 8).mul(255).to(torch.uint8).numpy(),
            mode="L",
        ).resize((120, 120), Image.Resampling.NEAREST)
        canvas.paste(image, ((index % 4) * 128 + 4, (index // 4) * 128 + 4))
    with torch.inference_mode():
        energies = MODEL(generated, labels)
    return canvas, {
        "digit": int(label),
        "langevin_steps": int(steps),
        "mean_energy": float(energies.mean()),
        "energy_standard_deviation": float(energies.std()),
    }


with gr.Blocks(title="Energy Pocket") as demo:
    gr.Markdown(
        "# Energy Pocket\n"
        "Watch random pixels descend a learned class-conditional energy landscape "
        "through stochastic Langevin dynamics."
    )
    with gr.Row():
        label = gr.Slider(0, 9, value=2, step=1, label="Digit class")
        seed = gr.Slider(0, 100_000, value=2069, step=1, label="Seed")
        steps = gr.Slider(10, 160, value=40, step=10, label="Langevin steps")
    initial = sample_energy(2, 2069, 40)
    output = gr.Image(value=initial[0], label="Energy-minimized samples")
    metrics = gr.JSON(value=initial[1], label="Energy readout")
    button = gr.Button("Descend the energy landscape", variant="primary")
    button.click(sample_energy, inputs=[label, seed, steps], outputs=[output, metrics])


if __name__ == "__main__":
    demo.launch()