source/app.py

from __future__ import annotations

from pathlib import Path

import gradio as gr
import numpy as np
import torch
from model import ConditionalCodePrior, VectorQuantizedAutoencoder
from PIL import Image
from safetensors.torch import load_file

ARTIFACT_DIR = Path(__file__).resolve().parent / "artifacts" / "vq-pocket"
AUTOENCODER = VectorQuantizedAutoencoder()
AUTOENCODER.load_state_dict(load_file(ARTIFACT_DIR / "vq_vae.safetensors"))
AUTOENCODER.eval()
PRIOR = ConditionalCodePrior(codebook_size=AUTOENCODER.codebook_size)
PRIOR.load_state_dict(load_file(ARTIFACT_DIR / "code_prior.safetensors"))
PRIOR.eval()


@torch.inference_mode()
def generate_tokens(
    label: int,
    seed: int,
    temperature: float,
) -> tuple[Image.Image, Image.Image, dict]:
    labels = torch.full((12,), int(label), dtype=torch.long)
    codes = PRIOR.generate(labels, seed=int(seed), temperature=float(temperature))
    images = AUTOENCODER.decode_indices(codes)
    gallery = Image.new("L", (512, 384), color=0)
    for index, pixels in enumerate(images):
        image = Image.fromarray(
            pixels[0].mul(255).clamp(0, 255).to(torch.uint8).numpy(),
            mode="L",
        ).resize((120, 120), Image.Resampling.NEAREST)
        gallery.paste(image, ((index % 4) * 128 + 4, (index // 4) * 128 + 4))
    palette = np.array(
        [
            [int((code * 73) % 255), int((code * 151) % 255), int((code * 211) % 255)]
            for code in range(AUTOENCODER.codebook_size)
        ],
        dtype=np.uint8,
    )
    token_map = palette[codes[0].numpy()]
    token_image = Image.fromarray(token_map, mode="RGB").resize(
        (384, 384),
        Image.Resampling.NEAREST,
    )
    unique_sequences = len({row.numpy().tobytes() for row in codes.flatten(1)})
    metadata = {
        "digit": int(label),
        "tokens_per_image": 16,
        "codebook_size": AUTOENCODER.codebook_size,
        "unique_sequences_in_batch": unique_sequences,
        "temperature": float(temperature),
    }
    return gallery, token_image, metadata


with gr.Blocks(title="VQ-Pocket") as demo:
    gr.Markdown(
        "# VQ-Pocket\n"
        "Generate digits as discrete 4x4 visual-token sequences, then decode the "
        "tokens through a compact VQ-VAE."
    )
    with gr.Row():
        label = gr.Slider(0, 9, value=4, step=1, label="Digit class")
        seed = gr.Slider(0, 100_000, value=2053, step=1, label="Sampling seed")
        temperature = gr.Slider(
            0.1, 1.5, value=0.85, step=0.05, label="Token temperature"
        )
    initial = generate_tokens(4, 2053, 0.85)
    with gr.Row():
        gallery = gr.Image(value=initial[0], label="Decoded generations")
        tokens = gr.Image(value=initial[1], label="First sample's 4x4 token map")
    metrics = gr.JSON(value=initial[2], label="Discrete-latent readout")
    button = gr.Button("Sample token sequences", variant="primary")
    button.click(
        generate_tokens,
        inputs=[label, seed, temperature],
        outputs=[gallery, tokens, metrics],
    )


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