/AlexNet-Experiments

Experiments with AlexNet architecture for Image Classification

Primary LanguagePython

AlexNet Experiments

This repository contains a PyTorch implementation of the AlexNet architeture described in ImageNet Classification with Deep Convolutional Neural Networks. The following experiments have been performed on the architecture:

  • Activation Unit: ReLU vs Tanh
  • Dropout v/s No Dropout
  • Overlapping Pooling v/s Non-Overlapping Pooling
  • Optimization Techniques: SGD v/s SGD with momentum v/s SGD with momentum and weight decay v/s ADAM.

The experiments were done on a 35 class subset of the ImageNet Dataset that can be downloaded from here.

Organisation

The code in the repository are organised as follows:

  • network.py: AlexNet arch. implementation
  • dataset.py: custom Pytorch Dataset implementation
  • train.py: training/validation/testing code
  • main.py: driver code

Requirements

The dataset may be downloaded from here.

Usage

Use the following to get help with usage.

python main.py --help

Results

The results of all the experiments can are compiled in results.pdf. The best model that attained over 76% accuracy on the the test-set can be downloaded from here.