Theano: configuring systems GPU for Theano crashes

[pranitapradhan91]

Hello,

I am very new to theano and working on gpu.
However, I have managed to follow the instructions given here https://lepisma.github.io/2015/07/30/up-with-theano-and-cuda/. Attached is my theanorc file .theanorc.txt.
When I 'import theano' it crashes and I get something like this. (see the picture attached)

My versions:
Theano 1.0.0rc1
pygpu 0.6.9
Not able to install pycuda (raise SSLError(e, request=request)
requests.exceptions.SSLError: HTTPSConnectionPool(host='www.lfd.uci.edu', port=443): Max retries exceeded with url: /~gohlke/pythonlibs/ (Caused by SSLError(SSLError("bad handshake: SysCallError(-1, 'Unexpected EOF')",),))

I am working on windows 10.

Your help is appreciated.

Regards
Pranita

[notoraptor]

Hi. First of all, there are more up-to-date installations instructions on Theano website : http://deeplearning.net/software/theano_versions/dev/install_windows.html

(you should not need to install neither pycuda nor Visual Studio with the latest theano version)

Also, 2 other remarks:

• To get a compiler, the recommended conda package is m2w64-toolchain instead of mingw
• pygpu 0.6.9 is too old and should not work with theano 1.0.0rc1. The latest pygpu version is 0.7.4 and can be installed via conda channel mila-udem with following command:

conda install -c mila-udem pygpu

Could you retry to install with updated instructions and then check if you still have problems ?

[pranitapradhan91]

Thank you for your quick help.

Below I have explained the ways I tried from the link you sent and also your suggestions.

1. First thing tried (without pygpu)
   PACKAGES VERSION
   python 2.7.14
   numpy 1.12.0
   scipy 1.0.0
   m2w64-toolchain 5.3.0
   mkl-service 1.1.2
   Theano 1.0.0rc1

import theano

error--- no module numpy

1. Second thing (without pygpu and updated numpy)
   PACKAGES VERSION
   python 2.7.14
   numpy 1.13.0
   scipy 1.0.0
   m2w64-toolchain 5.3.0
   mkl-service 1.1.2
   Theano 1.0.0rc1

import theano
WARNING (theano.tensor.blas): Using NumPy C-API based implementation for BLAS functions.

But does not work with gpu yet.

1. Third thing (installed pygpu 0.7.4, this updated numpy to 1.13.3)
   PACKAGES VERSION
   python 2.7.14
   numpy 1.13.3
   scipy 1.0.0
   m2w64-toolchain 5.3.0
   mkl-service 1.1.2
   Theano 1.0.0rc1

import theano
RuntimeError: To use MKL 2018 with Theano you MUST set "MKL_THREADING_LAYER=GNU" in your environement.

How can I do this?

Fourth thing (installed Theano==0.9.0 with conda)
PACKAGES VERSION
python 2.7.14
numpy 1.13.3
scipy 1.0.0
m2w64-toolchain 5.3.0
mkl-service 1.1.2
Theano 0.9.0

import theano
works without error, however does not use gpu. And when I run theano.test() I get following errors (.txt file attached).
errors.txt

*Everything was done in a virtual environment and .theanorc.txt file was same as in my earlier comment and was kept in home directory.

Is there anything I am still failing to do?
Waiting for suggestions. Thank you in advance.

Best,
Pranita

[notoraptor]

You should go with the third thing:

Third thing (installed pygpu 0.7.4, this updated numpy to 1.13.3)
PACKAGES VERSION
python 2.7.14
numpy 1.13.3
scipy 1.0.0
m2w64-toolchain 5.3.0
mkl-service 1.1.2
Theano 1.0.0rc1

About the error:

import theano
RuntimeError: To use MKL 2018 with Theano you MUST set "MKL_THREADING_LAYER=GNU" in your environement.

It is related to a bug recently found with latest versions of mkl package. You just need to set the environment variable as suggested by the error message. On a command line:

set MKL_THREADING_LAYER=GNU

Or via Windows Control Panel (set system environment variables), create a system variable named MKL_THREADING_LAYER with value GNU.

Also, to run theano.tests(), you may need to install package parameterized via pip:

pip install parameterized

[notoraptor]

NB: From your theanorc:

[global]
floatX = float32
device = cuda
mode=FAST_RUN

[cuda]
root = C:\Program Files\NVIDIA Corporation\Installer2\CUDAToolkit_9.0.{AE85B8EA-C063-4E5E-B333-151BFFB28009} 


[nvcc]
flags = -LC:\Users\si62qit\AppData\Local\Continuum\Anaconda2\libs
compiler_bindir = C:\Program Files (x86)\Microsoft Visual Studio 14.0\VC\bin
fastmath = True

If CUDA is correctly installed, you should not need [cuda] and [nvcc] sections. Consider having just this code into your theanorc:

[global]
floatX = float32
device = cuda

[pranitapradhan91]

Thank you.
This worked. But now I face another problem which is similar to the issue
https://github.com/Theano/Theano/issues/5768

Therefore I changed my .theanorc file
.theanorc.txt

Any suggestions? Waiting for reply.

Regards
Pranita

[notoraptor]

Have you already downloaded and installed cuDNN in any directory ? If not, it seems you need it.

If so, you may need to tell Theano where to find cuDNN libraries and headers, with config variable dnn.base_path. In theanorc, add these lines:

[dnn]
base_path=C:\Users\notoraptor\data\libraries\cudnn-8.0-windows10-x64-v7.0.3\cuda

(replace value of base_path with path to folder when you have installed cuDNN).

[pranitapradhan91]

Thank you for the reply.
I installed cuDNN. So basically here is what I did.

1. cudnn64_7.dll file is kept in C:Program FilesNVIDIA GPU Computing ToolkitCUDAv9.0bin
2. cudnn.h is kept in C:Program FilesNVIDIA GPU Computing ToolkitCUDAv9.0include
3. cudnn.lib is kept in C:Program FilesNVIDIA GPU Computing ToolkitCUDAv9.0libx64

Is this correct? Also .theanorc file remained same. As I included cuDNN in the toolkit folder as shown above.
.theanorc.txt

And now it runs.

Thank you very much for your help.

[notoraptor]

Yes, it should be correct, now. But your cuda root seems quite strange in your theanorc (C:\Program Files\NVIDIA Corporation\Installer2\CUDAToolkit_9.0.{AE85B8EA-C063-4E5E-B333-151BFFB28009}).

If you encounter some problems in the future, consider simplifying the theanorc file as I suggested above.

[wqp89324]

@notoraptor
In a virtual environment created by conda, I have almost exact versions of these libraries as in the previous post:

python 2.7.14
numpy 1.13.3
scipy 1.0.0
m2w64-toolchain 5.3.0
mkl-service 1.1.2
Theano 1.0.1
pygpu 0.7.5

However I still have the warning message: Using NumPy C-API based implementation for BLAS functions.
At this stage, I did not use a theanorc file yet. I'm on Windows.

Any idea how to fix my issue?
Thanks!