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.
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 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.
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.
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 .
.