Awesome

Artifical neural networks, anytime

This is a code repository to efficiently train a deconvolutional neural network with rectified linear units.

This code uses theano, pylearn2, cuda-convnet and is heavily based on Sander Dieleman's kaggle galaxy repo.

It also currently relies on a change to pylearn2.sandbox.cuda_convnet.pool.py that defines a grad method for the MaxPoolGrad class, which can be useful in some cases...

ICLR 2015 Paper Repo

If you came here via our paper An Analysis of Unsupervised Pre-training in Light of Recent Advances, please go here to access the experiments we ran.

Layout of the code

There are currently 3 main modules:

• datasets - generic dataset classes
• layers - layer definitions
• util - utilities for training, evaluating, and saving/loading checkpoints