More notes for myselfâ€¦ so it may not be helpful for you who bumped into here. ðŸ˜‰

Setting Theano correctly is not enough to ensure you can run deep learning software correctly. In our case, it will be Keras, and it can slow to a crawl if not setup properly.

Again, there could be many causes but I try to outline a clean step what I did, the performance I run a good setup, so you can compare. Hopefully you can glean some places where you did wrong.

My server has the following specifications finished running the steps outlined here.

OS: Ubuntu 14.04 LTS, X64
GPU: Nvidia Geforce GTX 780
Ubuntu 14.04 LTS
CUDA 7.5
Theano 0.7.0
Numpy 1.8.2
Kera 0.2.0
Scipy 0.13.3
NVIDIA-SMI 352.39
Graphics Driver Version: 352.39

Make sure your Theano configuration file, located at ~/.theanorc, is correct:

# Run this python code
from theano.sandbox.cuda.dnn import *
print(dnn_available())
print(dnn_available.msg)

Install CuDNN if you are using ConvNet. The basic implementations of convolution in Theano are significantly slower.

I also captured the environment variables, replace echio with your username:

Googling the error message doesnâ€™t help too much.

You need to add the location of the 3. into CPATH, LD_LIBRARY_PATH and LIBRARY_PATH. This is what my .bashrc looks like this (replace echio with your username):

If you instead see this error message:

You probably didnâ€™t have CUDA environment variables setup properly. See the above ~/.bashrc lines for correct setup.

Run this code again.

Run this python code

from theano.sandbox.cuda.dnn import * print(dnn_available()) print(dnn_available.msg)

You should see below when executed in a Python REPL.

This is good! Re-run your Keras code and hopefully it will be fast this timeâ€¦

This may or may not solve your problem, but it certainly solved some of my problems. You will probably have to learn to debug things a bit to figure out how to get it to run well.

Why is Keras Running So Slow?

未知

Run this python code

References