Skip to content

Latest commit

 

History

History
68 lines (51 loc) · 2.18 KB

README.md

File metadata and controls

68 lines (51 loc) · 2.18 KB

Neural Nets

Contents

About

This repo contains some personal implementations of neural networks done for practice implementing backprop and optimization from scratch (using only numpy). Currently all networks are sequential.

To install as python package, run:

python setup.py install

To compile Cython modules for testing run:

python setup.py build_ext --inplace

Branch 'Policies'

For development, push to the develop branch, then merge with master when ready.

Repository Contents

The neural_nets_dsr package contains the following subpackages and modules:

  • The activations package contains activation functions for network layers, along with an ActivationFunc class in the base.py module for creating new ones.

  • The cost_functions package contains several cost functions that can be used to train networks, and it also contains a base.py module with a class that allows the creation of new cost functions.

  • The layers package contains several layer implementations that can be used to construct networks.

  • The optim package contains optimization algorithms for training.

  • The network.py module contains the class that represents a network.

  • Finally, the utils.py module is for miscellaneous utility functions and classes.

Organization

Guidelines

  • Every activation and cost function should 'know' how to compute its own gradient.
  • Each layer should know how to forward and back propagate through itself.
  • Every optimizer should know how to perform updates on weights and biases.

Versioning

  • Tiny version change 0.0.x: Bugfix or minor change in implementation.
  • Minor version change 0.x.0: New feature added, but still compatible with previous versions.
  • Major version change x.0.0: Major refactoring or changes that break compatibility with previous versions.

Current Issues

  • Doubts about batchnorm derivative computation.
  • Numerical stability.

Back to top