ResNet18 for MNIST Classification

Overview

This project implements a Convolutional Neural Network (CNN) using the ResNet18 architecture for digit recognition on the MNIST dataset. The model leverages residual connections to enhance training efficiency and prevent gradient vanishing issues.

Features

• Utilizes ResNet18 with residual blocks.
• Supports MNIST dataset for digit classification.
• Implements data preprocessing with resizing, normalization, and grayscale-to-RGB conversion.
• Saves and loads trained model weights for reuse.
• Provides visualization of predictions with matplotlib.

Dataset

• MNIST Dataset:
  • 60,000 training images and 10,000 test images of handwritten digits (0–9).
  • Preprocessed by resizing images to 32x32 and converting grayscale to RGB (3 channels).

Model Architecture

• Residual Block: Includes skip connections for easier gradient flow.
• ResNet18 Layers:
  • Initial convolutional layer with ReLU activation.
  • 4 residual layers with increasing channels (64, 128, 256, 512).
  • Average pooling and fully connected output layer.

Requirements

• Python 3.x
• PyTorch
• torchvision
• numpy
• matplotlib

Install dependencies:

pip install torch torchvision numpy matplotlib

Training Parameters

• Epochs: 10
• Batch Size: 128
• Learning Rate: 0.01
• Momentum: 0.9
• Weight Decay: 5e-4

Usage

1. Clone the repository:

   git clone https://github.com/ZamoRzgar/ResNet18.git
   cd ResNet18

2. Train the model:

   python train.py

3. Evaluate the model:

   python test.py

4. Predict on sample images:

   python predict.py

Results

• Achieved 99.52% accuracy on MNIST test dataset.
• Visualized predictions for sample digits.

Observations

• Residual connections improve learning and prevent gradient vanishing.
• Larger input size (32x32) enhances performance with ResNet.
• Performance may vary with custom datasets or additional augmentations.

Future Enhancements

• Test model on more diverse datasets.
• Incorporate data augmentation techniques.
• Expand to deeper networks for complex image datasets.

References

1. He, Kaiming, et al. "Deep residual learning for image recognition." Proceedings of the IEEE conference on computer vision and pattern recognition. 2016.
2. MNIST dataset: http://yann.lecun.com/exdb/mnist/
3. PyTorch Documentation: https://pytorch.org

---

Author: Zamo Rzgar
Contact: [email protected]