README.md

nn-blocks

Contents

• Introduction
• Notebooks
• Documentation

Introduction

A neural network library built from scratch, without dedicated deep learning packages. Training and testing deep neural networks and utilizing deep learning best practices for multi-class classification with fully connected neural networks, text generation with recurrent neural networks, and regression with fully connected networks.

Notebooks

• 
  • image classification on the CIFAR-10 dataset
  • one-layer networks with Hinge and cross entropy losses
  • cyclical learning rate schedule for improved learning
  • exploring the effects of the initial learning rate of the cyclical learning rate schedule and L2 regularization strength on model performance, without hyperparameter search
• 
  • image classification on the CIFAR-10 dataset
  • two-layer networks with cross entropy loss
  • cyclical learning rate schedule for improved learning
  • Xavier initialization for avoiding activation saturation
  • Bayesian hyperparameter search with hyperopt
• 
  • image classification on the CIFAR-10 dataset
  • k-layer networks with cross-entropy loss
  • cyclical learning rate schedule for improved learning
  • Xavier initialization for avoiding activation saturation
  • Bayesian hyperparameter search with hyperopt
  • dropout and batch normalization for avoiding overfitting
  • data augmentation for avoiding overfitting with imgaug
  • AdaGrad for more efficient gradient descent optimization
• 
  • generating text from Harry Potter books and Donald Trump tweets with RNNs
  • one-hot encoding, gradient clipping and smoothed loss, etc.
• 
  • linear and non-linear regression

Documentation

The documentation, which was built with Sphinx, is hosted here.

To generate it locally, do the following:

cd nn-blocks
conda env create -f environment.yml
conda activate nn_blocks_env
cd docs
make clean
make html
google-chrome _build/html/index.html