A CUDNN minimal deep learning training code sample using LeNet.

Original readme: https://github.com/tbennun/cudnn-training

This fork enables the cmake option USE_OPENCL, to build on OpenCL 1.2

To build

• Install latest coriander
• Then do:

source ~/coriander/activate
git clone https://github.com/hughperkins/cudnn-training
cd cudnn-training
mkdir build
cd build
ccmake ..
# set `USE_CUDA` to `OFF`, and `USE_OPENCL` to `ON`
# press 'c' then 'g'
make
./lenet

Running

Extract the MNIST training and test set files (*-ubyte) to a directory (if gflags are not used, the default is the current path).

You can also use the pre-trained weights published along with CUDNN, using the "pretrained" flag.