[PYTHON] A story about how theano was moved with TSUBAME 2.0

abstract It is a record until theano is run on TSUBAME 2.0.

In TSUBAME 2.0, you need to build all python related items by yourself. I installed numpy, scipy, theano by referring to this site, but scipy.test () failed. And theano didn't work either.

2016/7/25 postscript It seems that this site has disappeared, so postscript. First, register the public key on the portal. Set ssh with

~/.ssh/config

Host login-t2.g.gsic.titech.ac.jp                                            
    HostName login-t2.g.gsic.titech.ac.jp                                      
    User USER_ID                                                              
    IdentityFile WHERE_OF_PUBLIC_KEY                                                 
    RemoteForward 3128 proxy.noc.titech.ac.jp:3128

On the TSUBAME side,

~/.bashrc

export http_proxy="http://localhost:3128"                                      
export https_proxy="http://localhost:3128"                                     
export ftp_proxy="http://localhost:3128"

End of postscript

As a result of trial and error, installing numpy using OpenBLAS forced theano to work properly.

[Mutism]

• Compiler: gcc
• python-2.7.7
• openblas-0.2.13
• numpy-1.8.1
• scipy-0.14.0
• theano-0.7.0

[Demerit]

• I'm losing because it's probably faster to use mkl than openblas

First, go to 3.4 (install matplotlib) on the above site.

OpenBLAS

cd OpenBLAS
make BINARY=64 USE_THREAD=1 FC=gfortran
make PREFIX=~/.local/OpenBLAS/ install 

Put it in the PATH

export LIBRARY_PATH=~/.local/lib:~/.local/OpenBLAS/lib:$LIBRARY_PATH
export LD_LIBRARY_PATH=~/.local/lib:~/.local/OpenBLAS/lib:$LD_LIBRARY_PATH
export PATH=~/.local/bin:~/.local/OpenBLAS/bin:$PATH
export BLAS=/home/....../.local/OpenBLAS/lib/libopenblas.a
export LAPACK=/home/....../.local/OpenBLAS/lib/libopenblas.a

numpy

cd numpy
cp site.cfg.example site.cfg

Here, and [here](http://osdf.github.io/blog/numpyscipy-with-openblas- For-ubuntu-1204-second-try.html), write site.cfg as follows (Hereafter, please rewrite PATH as appropriate)

site.cfg

[default]
library_dirs = /home/....../.local/OpenBLAS/lib

[openblas]
libraries = openblas
library_dirs = /home/....../.local/OpenBLAS/lib
include_dirs = /home/....../.local/OpenBLAS/include

[atlas]
atlas_libs = openblas
library_dirs = /home/....../.local/OpenBLAS/lib

[lapack]
lapack_libs = openblas
library_dirs = /home/....../.local/OpenBLAS/lib

Somehow, if you don't set atlas and lapack, segfault will occur when testing the dot () function with scipy.

Check if it was set properly

python setup.py config

If openblas, atlas, lapack are displayed as FOUND, move on to installation

python setup.py build
python setup.py install
cd ~
python -c "import numpy;numpy.test(verbose=10)"

scipy There seems to be no need to do anything about scipy

cd scipy
python setup.py build
python setup.py install
cd ~
python -c "import scipy;scipy.test(verbose=10)"

Theano

cd theano
python setup.py develop --prefix=~/.local
python setup.py install

Create ~ / .theanorc as follows

.theanorc

[global]
floatX = float32
device = gpu
openmp = True


[nvcc]
fastmath = True

[blas]
ldflags =  -lopenblas

[cuda]
root = /usr/apps.sp3/cuda/6.0/
[gcc]
cxxflags = -I/usr/apps.sp3/mpi/openmpi/1.8.2/g4.3.4_cuda6.0/include -I/usr/apps.sp3/mpi/openmpi/1.8.2/g4.3.4_cuda6.0/include/openmpi -L/usr/apps.sp3/mpi/openmpi/1.8.2/g4.3.4_cuda6.0/lib -lmpi -lmpi_cxx

Then run your favorite program. If the sample of Deep Learning Tutorials works, I wonder if it can be managed. ..