如何在Ubuntu 16.04上使用Theano和GPU？

Question

I use the following script to test if GPU is working:

#!/usr/bin/env python
from theano import function, config, shared, sandbox
import theano.tensor as T
import numpy
import time

vlen = 10 * 30 * 768  # 10 x #cores x # threads per core
iters = 1000

rng = numpy.random.RandomState(22)
x = shared(numpy.asarray(rng.rand(vlen), config.floatX))
f = function([], T.exp(x))
print f.maker.fgraph.toposort()
t0 = time.time()
for i in xrange(iters):
    r = f()
t1 = time.time()
print 'Looping %d times took' % iters, t1 - t0, 'seconds'
print 'Result is', r
if numpy.any([isinstance(x.op, T.Elemwise) for x in f.maker.fgraph.toposort()]):
    print('Used the cpu')
else:
    print('Used the gpu')

When I run it, I get:

http://pastebin.com/wM9jaGMF

The interesting part is at the end:

ERROR (theano.sandbox.cuda): Failed to compile cuda_ndarray.cu: ('nvcc return status', 1, 'for cmd', 'nvcc -shared -O3 -m64 -Xcompiler -DCUDA_NDARRAY_CUH=c72d035fdf91890f3b36710688069b2e,-DNPY_NO_DEPRECATED_API=NPY_1_7_API_VERSION,-fPIC,-fvisibility=hidden -Xlinker -rpath,/home/moose/.theano/compiledir_Linux-4.4--generic-x86_64-with-Ubuntu-16.04-xenial-x86_64-2.7.11+-64/cuda_ndarray -I/home/moose/.local/lib/python2.7/site-packages/theano/sandbox/cuda -I/usr/lib/python2.7/dist-packages/numpy/core/include -I/usr/include/python2.7 -I/home/moose/.local/lib/python2.7/site-packages/theano/gof -o /home/moose/.theano/compiledir_Linux-4.4--generic-x86_64-with-Ubuntu-16.04-xenial-x86_64-2.7.11+-64/cuda_ndarray/cuda_ndarray.so mod.cu -L/usr/lib -lcublas -lpython2.7 -lcudart')
WARNING (theano.sandbox.cuda): CUDA is installed, but device gpu is not available  (error: cuda unavailable)

My system

• I use Ubuntu 16.04.
• I've installed CUDA through the standard repos ( V7.5.17 ). nvcc --version works.
• I've installed Theano via pip
• I have CuDNN 4 (works with TensorFlow)
• I set CUDA_ROOT=/usr/bin/ and LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu/ (I'm not sure if that is correct)

My ~/.theanorc is

[global]
exception_verbosity=high
device=gpu
floatX=float32

[cuda]
root=/usr/bin/

Paths

I think the installation from the standard repos might make things different from a manual installation. Here are some paths which might uncover some problems:

/usr/bin/nvcc
/usr/lib/x86_64-linux-gnu/libcuda.so
/usr/lib/x86_64-linux-gnu/libcudart.so
/usr/lib/nvidia-cuda-toolkit
/usr/include/cudnn.h

Question

How can I make it work?

Answer 1

I'm not exactly sure what solved the issue, but one or both of the following ( source )

sudo apt-get install python-numpy python-scipy python-dev python-pip python-nose g++ libopenblas-dev libblas-dev git
echo -e "\n[nvcc]\nflags=-D_FORCE_INLINES\n" >> ~/.theanorc

Answer 2

I am writing a more general answer here in case others find themselves in a similar situation. First, see that theano dependencies are installed, as explained here . you should install nvidia driver as described here , use sudo ubuntu-drivers devices to determine which driver is recommended and install it with sudo apt-get install nvidia-xxx (xxx = 375 at this time). Then see that the nvidia driver is being used by opening "additional drivers" window (from terminal software-properties-gtk --open-tab=4 ). Setup a ~/.theanorc text file as follows:

[global]
exception_verbosity=high
device=gpu
floatX=float32

[cuda]
root=/usr/bin/

[nvcc]
flags=-D_FORCE_INLINES

[lib]
cnmem = 1

[lib] section is not necessary but on my laptop performance was about 2 times faster when adding it to .theanorc.