SHARKonvnet

GPU CNN's for SHARK.

Installation

• Cuda Toolkit https://developer.nvidia.com/cuda-downloads
• cuDNN https://developer.nvidia.com/cudnn
• SHARK http://image.diku.dk/shark/

Makefile

• Here be hints on how to make it compile.

MNIST

• Change directory to scripts and run python prep_mnist_data.py. This will download, unpack and normalize the MNIST dataset to data/mnist/.
• Change directory to the root of the project and run make mnist. This will compile tests/MnistTest.cpp into obj/MnistTest.o.
• Run with ./obj/MnistTest.o. You should see some info about the convnet (specified in tests/MnistTest.cpp), and then some progress info every 500 itertation. After 2000 iterations the 01-loss should be around 4%. It converges to a little below 1%.