January 25, 2014

Deep learning libraries in python

Theano

Theano is a generic mathematical expression evaluation library. The main advantage is, it compiles to either CPU or GPU. By default it uses cpu. 

Pylearn2

Pylearn2 is based on theano. So you can get advantage of running on GPU.

Download from https://github.com/lisa-lab/pylearn2

Models
  • Logistic Regression 
  • K-means 
  • Multilayer Perceptron (MLP)
  • Restricted Boltzmann Machines (RBM)
  • Deep Boltzmann Machines (DBM)
  • Ising 
  • Auto Encoder: Autoencoders, denoising autoencoders, and stacked DAEs.
  • Maxout Networks (Maxout)
It is not limited to above models and any one can add new model. Above models are pre implemented.

Hebel

It is also GPU accelerated deep learning library. Its based on PyCuda.

Models
  • Logistic Regression
  • Neural network regression
  • Muti-Task neural net
Download from https://github.com/hannes-brt/hebel

Deepnet

GPU accelerated. Based on cudamat. 

Models
  • Feed-forward Neural Nets
  • Restricted Boltzmann Machines
  • Deep Belief Nets
  • Autoencoders
  • Deep Boltzmann Machines
  • Convolutional Neural Nets

Others

Python tutorial on Restricted Bolzmann Machines - https://github.com/echen/restricted-boltzmann-machines
Modular Restricted Bolzmann machine implementation. Its based on theano. https://github.com/benanne/morb
Matrix Library for cuda - https://github.com/deeplearningais/CUV

No comments:

Post a Comment