MNIST-Image-Classifier

This repository contains a notebook for a PyTorch-based neural network model designed to classify images using the MNIST handwritten images dataset.

Data source from Kaggle: MNIST Dataset on Kaggle

Neural Network Architecture

Overview

The network architecture is as follows:

• Input Layer: Flattened 28x28 pixel images.
• Three fully connected (linear) layers with ReLU activation.
• Output Layer: 10 classes for classification.

Network Details

• Layer 1 (Input Layer):

  • Input Size: 784 (28x28 pixels)
  • Output Size: 100
  • Activation Function: ReLU

• Layer 2:

  • Input Size: 100
  • Output Size: 50
  • Activation Function: ReLU

• Layer 3 (Output Layer):

  • Input Size: 50
  • Output Size: 10 (for classification)

Forward Pass

The forward pass of this network involves the following steps:

1. Input data is flattened from 28x28 to 784 dimensions.
2. The first linear layer applies a ReLU activation.
3. The second linear layer applies another ReLU activation.
4. The third linear layer (output layer) produces raw scores/logits for classification.

The final output is a 1D tensor, suitable for classification tasks.

Results & Training

• Training was completed for 20 epochs on a dataset of 60,000 images and a batch size of 5.
• Average loss for each epoch was calculated using cross-entropy loss function.

• Sample testing images are shown below with the model predictions.