https://github.com/krishna-smapca/neural-network-using-pytorch

This repo contains the implemention of few neural network on Fashion Mnist Dataset, designed for Multi-Class classification
https://github.com/krishna-smapca/neural-network-using-pytorch

ai ann artificial-intelligence classification deep-neural-networks machine-learning neural-networks python3 pytorch pytorch-implementation

Last synced: 3 days ago
JSON representation

This repo contains the implemention of few neural network on Fashion Mnist Dataset, designed for Multi-Class classification

• Host: GitHub
• URL: https://github.com/krishna-smapca/neural-network-using-pytorch
• Owner: Krishna-Smapca
• Created: 2025-02-07T07:10:09.000Z (7 days ago)
• Default Branch: main
• Last Pushed: 2025-02-07T12:08:37.000Z (7 days ago)
• Last Synced: 2025-02-07T12:22:32.337Z (7 days ago)
• Topics: ai, ann, artificial-intelligence, classification, deep-neural-networks, machine-learning, neural-networks, python3, pytorch, pytorch-implementation
• Language: Jupyter Notebook
• Homepage:
• Size: 14.6 KB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# Neural-Network-using-Pytorch

This repository contains three different models for multi-class classification using PyTorch, all trained on the Fashion MNIST dataset. The models include an Artificial Neural Network (ANN), a custom Convolutional Neural Network (CNN), and a CNN model with pre-trained ResNet weights. The goal of the project is to practice neural network design and enhance model performance using various techniques.

## Models

### 1. **Artificial Neural Network (ANN)**

   - **Input size**: 784 (28x28 flattened pixels from Fashion MNIST)

   - **Architecture**: 

     - Input → Linear (784 → 128) → BatchNorm → ReLU → Dropout

     - Linear (128 → 64) → BatchNorm → ReLU → Dropout

     - Output → Linear (64 → 10)

   - **Features**:

     - Batch normalization and dropout to speed up learning and reduce overfitting.

     - Predicts 10 classes (Fashion MNIST categories).

  

### 2. **Custom Convolutional Neural Network (CNN)**

   - **Input size**: 28x28 image (no flattening required)

   - **Architecture**: 

     - Feature Extraction: 2 Convolutional layers with ReLU activation and BatchNorm

     - Max Pooling layers

     - Fully connected classifier with Dropout for regularization

   - **Features**:

     - 1st Convolution Layer: (1 → 32 channels)

     - 2nd Convolution Layer: (32 → 64 channels)

     - Output layer with 10 classes (Fashion MNIST categories).

     - Batch normalization and dropout used for better learning and complexity reduction.

### 3. **CNN with Pre-trained ResNet Weights**

   - **Input size**: 28x28 image (no flattening required)

   - **Feature Extraction**: Pre-trained ResNet layers with frozen parameters

   - **Classification**: Custom fully connected layer (fine-tuned)

   - **Architecture**:

     - **Feature extraction**: Pre-trained ResNet (frozen layers)

     - **Classification**: Linear (2048 → 128) → ReLU → Dropout → Linear (128 → 64) → ReLU → Dropout → Linear (64 → 10)

   - **Features**:

     - Transfer learning with frozen feature extraction layers.

     - Fine-tuning of the classification layer to adapt to the Fashion MNIST dataset.

## Performance

- **ANN**: 

   - Training Accuracy: 99% 

   - Test Accuracy: 82%

   - Shows overfitting (higher training accuracy, lower test accuracy).

  

- **Custom CNN**:

   - Training Accuracy: 82%

   - Test Accuracy: 82%

  

- **CNN with Pre-trained ResNet**:

   - Training Accuracy: 87%

   - Test Accuracy: 87%

   - Better performance with higher epochs.

## Requirements

- PyTorch

- torchvision

- numpy

- matplotlib

## Installation

```bash

pip install torch torchvision numpy matplotlib

```

## Usage

You can easily train and evaluate each of the models by running the respective Python files. The dataset will be automatically downloaded if not already available.

---

This repository provides a clear comparison of different neural network architectures on the Fashion MNIST dataset, highlighting the advantages of CNNs and transfer learning over traditional ANN models.