My fork of Alex Krizhevsky's cuda-convnet from 2013 where I added dropout, among other features.
Cuda Python C++
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.

README.rst

This is my fork of the cuda-convnet convolutional neural network implementation written by Alex Krizhevsky.

cuda-convnet has quite extensive documentation itself. Find the MAIN DOCUMENTATION HERE.

Update: A newer version, cuda-convnet 2, has been released by Alex. This fork is still based on the original cuda-convnet.

Additional features

This document will only describe the small differences between cuda-convnet as hosted on Google Code and this version.

Dropout

Dropout is a relatively new regularization technique for neural networks. See the Improving neural networks by preventing co-adaptation of feature detectors and Improving Neural Networks with Dropout papers for details.

To set a dropout rate for one of our layers, we use the dropout parameter in our model's layer-params configuration file. For example, we could use dropout for the last layer in the CIFAR example by modifying the section for the fc10 layer to look like so:

[fc10]
epsW=0.001
epsB=0.002
# ...
dropout=0.5

In practice, you'll probably also want to double the number of outputs in that layer.

CURAND random seeding

An environment variable CONVNET_RANDOM_SEED, if set, will be used to set the CURAND library's random seed. This is important in order to get reproducable results.

Updated to work with CUDA via CMake

The build configuration and code has been updated to work with CUDA via CMake. Run cmake . and then make. If you have an alternative BLAS library just set it with for example cmake -DBLAS_LIBRARIES=/usr/lib/libcblas.so ..