Skip to content
Port of Alex Krizhevsky's cuda-convnet to windows x64 . Original project
C Python C++ Shell
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.


Port to windows x64 of Alex Krizhevsky's cuda-convnet. Original project and documentation at


New option --fix-nan=1 . You can use it if you have "get nan or inf" error, wich sometimes popping up on some lower end GPU's. This option add unary operation wich replace NaN values with zeros in gradients of convolutinal layers. This is an experimantal option, it's possible that it's not always help.


NVIDIA videocad with cuda compute capability 2.0 or better

Windows x64

Cuda SDK 4.2 (will not work for 5.0, for 5.0 port look to original project commentaries or Wyvernbai blog

Visual Studio 2008 or later with x64 support(x64 is not installed by defult in VS, may require installation update).

Python 2.7 64bit; Neither Python 3.*, nor Python compiled for 32 bits will not work

NumPy package for 64bit Python NumPy for window and 64bit python included in Anaconda package and WinPython package. Pythonxy will not work - it's only 32 bit


Install WinPython or Anaconda for both 64bit Python and NumPy


  1. Make shure that you have PYTHONPATH enviromental variable set only to your 64 bit Python folder. If it set to other folder(s) you should replace $(PYTHONPATH) in the project settings in C/C++ and linker to absolute directory of 64bit python.

  2. create folder in (NVIDIA SDK)\C\src folder and create repository there (for example E:\ProgramData\NVIDIA Corporation\NVIDIA GPU Computing SDK 4.2\C\src\convnet) and pull

  3. unpack win_convnet\ into win_convnet folder

  4. build project pyconvnet for x64 (preferably Release), it should produce pyconvnet.pyd (dll for python) in win_convnet folder

  5. run synthetic test from :

python --layer-def=./example-layers/layers.gc.cfg --layer-params=./example-layers/layer-params.gc.cfg --data-provider=dummy-cn-192 --check-grads=1

Depending on your videocard some tests may fail, but majority should almoste always pass.

  1. Download real data from and run training on them, as described at

While building this project I've followed exellent instructions by Yalong Bai(Wyvernbai)from and used his precompiled dlls, with some changes (cuda 4.2 instead of 5.2)

You can’t perform that action at this time.