app.py

# -*- coding: utf-8 -*-
"""app.ipynb

Automatically generated by Colab.

Original file is located at
    https://colab.research.google.com/drive/104LGaVaxm2qBgfuuNORSOHuzKCmgOVKm
"""

### 1. Imports and setup ###

import gradio as gr
import torch
import numpy as np

from model import create_diffusion_model
from timeit import default_timer as timer
from typing import Tuple

# ------------------------------
# 2. Model preparation
# ------------------------------

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

# Digit vocabulary
digit_words = ["zero","one","two","three","four","five","six","seven","eight","nine"]
word_to_idx = {w:i for i,w in enumerate(digit_words)}

# Create diffusion model
model = create_diffusion_model(
    emb_dim=64,
    timesteps=100,
    seed=42,
    device=device
)

# Load trained weights
model.load_state_dict(
    torch.load(
        "text_to_digit_diffusion_mnist.pth",
        map_location=device
    )
)

model.eval()

# ------------------------------
# 3. Diffusion scheduler utils
# ------------------------------

timesteps = 100
betas = torch.linspace(1e-4, 0.02, timesteps).to(device)
alphas = 1.0 - betas
alphas_cumprod = torch.cumprod(alphas, dim=0)

def extract(a, t, x_shape):
    return a.gather(-1, t).reshape(-1,1,1,1).expand(x_shape)

# ------------------------------
# 4. Predict / Generate function
# ------------------------------

@torch.inference_mode()
def predict(text: str) -> Tuple[np.ndarray, float]:
    start_time = timer()

    text = text.strip().lower()

    # Normalize input
    if text.isdigit():
        label = int(text)
    else:
        label = word_to_idx.get(text, None)

    if label is None or not (0 <= label <= 9):
        raise ValueError("Please enter a digit (0–9) or its word form.")

    labels = torch.tensor([label], device=device)
    uncond_labels = torch.tensor([0], device=device)

    # Start from noise
    x = torch.randn(1, 1, 28, 28, device=device)

    guidance_scale = 3.0

    for i in reversed(range(1, timesteps)):
        t = torch.full((1,), i, device=device)

        pred_cond = model(x, t, labels)
        pred_uncond = model(x, t, uncond_labels)

        pred_noise = pred_uncond + guidance_scale * (pred_cond - pred_uncond)

        beta_t = extract(betas, t, x.shape)
        alpha_t = extract(alphas, t, x.shape)
        alpha_bar_t = extract(alphas_cumprod, t, x.shape)

        pred_x0 = (x - torch.sqrt(1 - alpha_bar_t) * pred_noise) / torch.sqrt(alpha_bar_t)
        x = torch.sqrt(alpha_t) * pred_x0 + torch.sqrt(beta_t) * pred_noise

    img = (x.clamp(-1,1) + 1) / 2
    img = img[0,0].cpu().numpy()

    end_time = timer()
    gen_time = round(end_time - start_time, 4)

    return img, gen_time

# ------------------------------
# 5. Gradio app
# ------------------------------

title = "Text-to-Digit Diffusion (MNIST)"
description = (
    "A **conditional diffusion model** trained on MNIST. "
    "Type a digit (e.g. `seven` or `7`) to generate a handwritten number."
)
article = "Created by [Programming Ocean Academy](https://www.programming-ocean.com/)"

demo = gr.Interface(
    fn=predict,
    inputs=gr.Textbox(
        label="Enter digit (0–9 or word)",
        placeholder="seven or 7"
    ),
    outputs=[
        gr.Image(
            label="Generated Digit",
            type="numpy",
            width=256,
            height=256
        ),
        gr.Number(label="generation time (s)")
    ],
    title=title,
    description=description,
    article=article
)

# Launch demo
demo.launch(debug=False)