The objective of this post is guide you use Keras with CUDA on your Windows 10 PC.
Hardware: A graphic card from NVIDIA that support CUDA, of course.
Driver: Download and install the latest driver from NVIDIA or your OEM website
CUDA: Download and install version 7.5 from https://developer.nvidia.com/cuda-75-downloads-archive DO NOT install outdated driver intergrated in installer. Visual Studio integration and GPU Deployment Kit is not needed.
cuDNN: Download and copy all folder (there should be three: “bin”, “include”, “lib”) in zipped cuda folder to your CUDA installation folder (there should be “bin”, “include”, “lib”) https://developer.nvidia.com/cudnn You may need to register as an NVIDIA Accelerated Computing Developer Program member to process to download.
Python: Download and install Anaconda 2 from https://www.continuum.io/downloads Run
conda install mingw libpython after installation.
Visual Studio: Install version 2010, 2012, or 2013. Newer version is not supported. The free community version is fine. We just need the compiler.
Microsoft Visual C++ Compiler for Python 2.7: Download and install from http://www.microsoft.com/en-us/download/details.aspx?id=44266
GCC: Download and install TDM-GCC from http://tdm-gcc.tdragon.net/
Install Theano and Keras
Keras support Theano or Tensor Flow as backend. However, Tensor Flow with GPU is not support in Windows. So just use Theano as backend.
Just install as a common package of python
pip install theano keras
Write a plain text file named
.theanorc.txt if previous one is hard to create)in your user folder(C:\Users\<Your username here>\)
Contents is as following
floatX = float32
device = gpu
compiler_bindir=C:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\bin
enabled = True
“device = gpu” means use gpu resource
Replace “C:\Apps\Anaconda2” to your installation path of your Anaconda 2. There is no space between “flags=-L” and the path.
Replace “C:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\bin” to your installation path of Visual Studio. There should be a “cl.exe” at that path.
Replace “cnmem=0.75” to your free graphic card memory ratio. You may use GPU-Z from https://www.techpowerup.com/gpuz/ to determine how much memory can be allocated to Theano besides your normal usage (for screen display, etc.). Theano will eat that much memory (75% of memory in this configuration) when you initialize Theano every time.
Open python console and run
It may be a little slow, but eventually it will print something like
Using gpu device 0: <Your GPU> (CNMeM is enabled with initial size: <Your cnmem ratio> of memory, cuDNN <numbers, whatever>)
Stop python console if it looks fine.
To manually assign Theano backend, change following lines in
C:\Users\<Your username here>\.keras\keras.json
"image_dim_ordering": "tf", to
"image_dim_ordering": "th", to use Theano’s image channel order(BGR).
"backend": "tensorflow" to
"backend": "theano" to assign Theano as backend.
Just run the example from official repositry https://github.com/fchollet/keras/blob/master/examples/imdb_cnn.py
Wait and you will know if it works well. You may also watch GPU-Z to know how many GPU resource is been used.
Now you get a fully workable Keras instance with CUDA acceleration.