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