app.py

import gradio as gr
import numpy as np
import torch
import torch.nn as nn
from torchvision import models, transforms
from PIL import Image


# =========================
# 1. MODEL DEFINITION
# (must match training exactly)
# =========================
def get_densenet121_mnist():
    model = models.densenet121(weights=None)

    # Grayscale: 1 input channel (same as training)
    model.features.conv0 = nn.Conv2d(1, 64, kernel_size=7, stride=2, padding=3, bias=False)

    # 10 output classes (same as training)
    num_ftrs = model.classifier.in_features
    model.classifier = nn.Linear(num_ftrs, 10)

    return model


# =========================
# 2. LOAD MODEL
# =========================
def load_model():
    model = get_densenet121_mnist()
    state_dict = torch.load("best_densenet_mnist.pth", map_location="cpu")
    model.load_state_dict(state_dict, strict=True)  # strict=True since arch matches exactly
    model.eval()
    return model

model = load_model()


# =========================
# 3. PREPROCESS IMAGE
# (must match training transform exactly)
# =========================
preprocess = transforms.Compose([
    transforms.Grayscale(num_output_channels=1),
    transforms.Resize((32, 32)),        # same as training
    transforms.ToTensor(),
    transforms.Normalize((0.1307,), (0.3081,))  # same as training
])


# =========================
# 4. PREDICT
# =========================
LABELS = [str(i) for i in range(10)]

def predict(image):
    image = Image.fromarray(image).convert("L")  # ensure grayscale PIL
    x = preprocess(image).unsqueeze(0)           # (1, 1, 32, 32)

    with torch.no_grad():
        output = model(x)
        probs = torch.softmax(output, dim=1).squeeze()

    return {LABELS[i]: float(probs[i]) for i in range(10)}


# =========================
# 5. GRADIO UI
# =========================
demo = gr.Interface(
    fn=predict,
    inputs=gr.Image(type="numpy"),
    outputs=gr.Label(num_top_classes=3),   # shows top-3 predictions with confidence
    title="DenseNet121 MNIST Classifier",
    description="Upload a handwritten digit (0–9). Model: DenseNet121 trained on MNIST."
)

demo.launch()