🧠 TinyCNN for MNIST (94K params)

This repository contains a lightweight Convolutional Neural Network (CNN) designed for the MNIST handwritten digit classification task.
The model is optimized to be small, fast, and easy to deploy, suitable for both research and educational purposes.

📌 Model Summary

Attribute	Value
Model Name	TinyCNN
Dataset	MNIST (28×28 grayscale digits)
Total Parameters	~94,410
Architecture	Conv-BN-ReLU ×3 → Global Avg Pool → FC
Input Shape	(1, 28, 28)
Output Classes	10
Framework	PyTorch

🏗 Architecture Overview

Input: 1×28×28

Conv Block 1: Conv(1→32, 3×3) → BatchNorm → ReLU → MaxPool(2×2)

Conv Block 2: Conv(32→64, 3×3) → BatchNorm → ReLU → MaxPool(2×2)

Conv Block 3: Conv(64→128, 3×3) → BatchNorm → ReLU → MaxPool(2×2)

Global Average Pooling

Fully Connected Layer → 10 output classes

This architecture emphasizes parameter efficiency while maintaining strong representation capability.

⚙️ Installation

pip install torch torchvision

🚀 Load Model From Hub

import torch
from model import TinyCNN  # Ensure this file is included in your repo

model = TinyCNN(num_classes=10)

state_dict = torch.hub.load_state_dict_from_url(
    "https://huggingface.co/<your-username>/<your-model-repo>/resolve/main/tinycnn_mnist.pth"
)
model.load_state_dict(state_dict)
model.eval()

🖼 Example Inference

import torch
from torchvision import transforms
from PIL import Image

transform = transforms.Compose([
    transforms.Grayscale(),
    transforms.Resize((28, 28)),
    transforms.ToTensor()
])

img = Image.open("digit.png")
x = transform(img).unsqueeze(0)  # shape: (1, 1, 28, 28)

with torch.no_grad():
    logits = model(x)
    pred = logits.argmax(dim=1).item()

print("Predicted digit:", pred)