Theano (Deep learning Tool) uses GPU for accelerated configuration and use

Recently used Theano wrote the MLP and CNN program, because the training sample large, CPU speed so slow, and then found a computer with Naivid graphics card configuration using the GPU, encountered a lot of problems, recorded as follows:

Platform Description:

System: WindowsXP

python:2.7, it is recommended to use Python (x, y) directly, including the Theano required NumPy library, save your own configuration

theano:0.6

cuda:3.0

1 Downloads

Download Install Theano:easy_install Theano

Download Cuda: Download drivers, toolkits, software development kits from http://www.nvidia.cn/object/cuda_get_cn_old.html, install sequentially

Note: Students are best to download the 3.0 version of the first, my earliest version of the 2.3, but found that Theano Cuda module on the 2.3 support is very poor, changed. \theano\sandbox\cuda\. CU inside a lot of code, such as Cudamemgetinfor (for the 3.0 version of the interface, 2.3 for the cumemgetinfor, see more http://stackoverflow.com/ questions/7386990/cuda-identifier-cudamemgetinfo-is-undefined), the final compilation is correct, but there is a difficult to deal with the link bug. When you finally switch to 3.0, only a small portion of the code is changed to run successfully (see instructions later).

2 configuration

Open the Theano configuration file (typically in the root directory of the user folder, such as C:\Documents and Settings\administrator\, named. theanorc.txt) and add the following configuration

[NVCC]

Flags=-lc:\python27\libs

Compiler_bindir=c:\program Files\Microsoft Visual Studio 9.0\vc\bin

Note: Version 2.3 only supports MSVC8.0 and MSVC9.0.

Then import Theano.sandbox.cuda to perform the initial test, the following configuration after successful

[Global]

device = GPU

Floatx = float32

The last run theano/misc/check_blas.py to test, or directly import Theano; Theano.test ()

3 problems encountered

1 class "Cudadeviceprop" has no conponent of named "Concurrentkernels"

Follow up on Cuda_ndarray.cu, near line 2796

#if cudart_version >= 3000
Put_in_dict (Dict, "Concurrentkernels", deviceprop.concurrentkernels);
#endif

Find Cudadeviceprop class under C:\CUDA, see there is no concurrentkernels, although vision is 3.0

Solution: Comment out the above section

2 Fatel Error C1083: Cannot open include file: Stdint.h:no such files or directory

Workaround:

To Googlecode download Http://msinttypes.googlecode.com/files/msinttypes-r26.zip, extract will get three files, put Inttypes.h and stdint.h to VC's include directory on it.

I installed the VS2008, installed to the default location, so the include path is:
C:\Program Files\Microsoft Visual Studio 9.0\vc\include

3 How to view GPU status

Download Gpuz

At last, although this machine with Nvidia is very old and old, with GPU acceleration, it's faster than running on my Y470.

GPU, very good very powerful.

Theano (Deep learning Tool) uses GPU for accelerated configuration and use