/hipsternet

All the hipster things in Neural Net in a single repo

Primary LanguagePythonThe UnlicenseUnlicense

hipsternet

All the hipster things in Neural Net in a single repo: hipster optimization algorithms, hispter regularizations, everything!

Note, things will be added over time, so not all the hipsterest things will be here immediately. Also don't use this for your production code: use this to study and learn new things in the realm of Neural Net, Deep Net, Deep Learning, whatever.

What's in it?

Network Architectures

  1. Convolutional Net
  2. Feed Forward Net
  3. Recurrent Net
  4. LSTM Net
  5. GRU Net

Optimization algorithms

  1. SGD
  2. Momentum SGD
  3. Nesterov Momentum
  4. Adagrad
  5. RMSprop
  6. Adam

Loss functions

  1. Cross Entropy
  2. Hinge Loss
  3. Squared Loss
  4. L1 Regression
  5. L2 Regression

Regularization

  1. Dropout
  2. Your usual L1 and L2 regularization

Nonlinearities

  1. ReLU
  2. leaky ReLU
  3. sigmoid
  4. tanh

Hipster techniques

  1. BatchNorm
  2. Xavier weight initialization

Pooling

  1. Max pooling
  2. Average pooling

How to run this?

  1. Install miniconda http://conda.pydata.org/miniconda.html
  2. Do conda env create
  3. Enter the env source activate hipsternet
  4. [Optional] To install Tensorflow: chmod +x tensorflow.sh; ./tensorflow.sh
  5. Do things with the code if you want to
  6. To run the example:
  7. python run_mnist.py {ff|cnn}; cnn for convnet model, ff for the feed forward model
  8. python run_rnn.py {rnn|lstm|gru}; rnn for vanilla RNN model, lstm for LSTM net model, gru for GRU net model
  9. Just close the terminal if you done (or source deactivate, not a fan though)

What can I do with this?

Do anything you want. I licensed this with Unlicense License http://unlicense.org, as I need to take a break of using WTFPL license.