Skip to content

A CUDNN minimal deep learning training code sample using LeNet.

Notifications You must be signed in to change notification settings

geotyper/cudnn-training

 
 

Repository files navigation

A CUDNN minimal deep learning training code sample using LeNet.

Prerequisites

Set the CUDNN_PATH environment variable to where CUDNN is installed.

Compilation

The project can either be compiled with CMake (cross-platform) or Visual Studio.

To compile with CMake, run the following commands:

~: $ cd cudnn-training/
~/cudnn-training: $ mkdir build
~/cudnn-training: $ cd build/
~/cudnn-training/build: $ cmake ..
~/cudnn-training/build: $ make

If compiling under linux, make sure to either set the CUDNN_PATH environment variable to the path CUDNN is installed to, or extract CUDNN to the CUDA toolkit path.

To enable gflags support, uncomment the line in CMakeLists.txt. In the Visual Studio project, define the macro USE_GFLAGS.

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 load and save pre-trained weights (e.g., published along with CUDNN), using the "pretrained" and "save_data" flags respectively.

About

A CUDNN minimal deep learning training code sample using LeNet.

Resources

Stars

Watchers

Forks

Packages

No packages published

Languages

  • Cuda 88.9%
  • C++ 9.1%
  • CMake 2.0%