Can not import cupy with chainer 4.1.0

Question

I have a project in python2.7.11 using chainer and cupy. After updating version of Chainer from 1.22 to 4.1.0, I can not using cupy

>>> import cupy
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
ImportError: No module named cupy

This is the way I installed chainer1.22. It was able to import cupy successfully without installing cupy.

pip install chainer==1.22,

Check

>>> import chainer
>>> import cupy
>>> chainer.cuda.available
True

I can see that CUDA is available with chainer1.22

So I uninstall chainer 1.22

pip uninstall chainer

Install chainer 4.1.0

pip install chainer==4.1.0

Result is

>>> import chainer
>>> import cupy
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
ImportError: No module named cupy
>>> chainer.cuda.available
False

Can see that cuda is not available with chainer version 4.1.0. So I try to install cupy with pip

pip install cupy

Log when installing

Requirement already satisfied: numpy>=1.9.0 in c:\users\UserName\appdata\local\continuum\anaconda2\lib\site-packages (from cupy)
Requirement already satisfied: six>=1.9.0 in c:\users\UserName\appdata\local\continuum\anaconda2\lib\site-packages (from cupy)
Requirement already satisfied: fastrlock>=0.3 in c:\users\UserName\appdata\local\continuum\anaconda2\lib\site-packages (from cupy)
Building wheels for collected packages: cupy
  Running setup.py bdist_wheel for cupy ... /
.
.
.
Microsoft (R) C/C++ Optimizing Compiler Version 15.00.30729.01 for x64
  Copyright (C) Microsoft Corporation.  All rights reserved.

  tmpxft_000027f8_00000000-1.cpp
  nvcc fatal   : Host compiler targets unsupported OS.
  error: command 'C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v8.0\\bin/nvcc.exe' failed with exit status 1

  ----------------------------------------
  Failed building wheel for cupy
  Running setup.py clean for cupy
Failed to build cupy
Installing collected packages: cupy
.
.
.
Microsoft (R) C/C++ Optimizing Compiler Version 15.00.30729.01 for x64
    Copyright (C) Microsoft Corporation.  All rights reserved.

    tmpxft_0000456c_00000000-1.cpp
    nvcc fatal   : Host compiler targets unsupported OS.
    error: command 'C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v8.0\\bin/nvcc.exe' failed with exit status 1

    ----------------------------------------
Command "C:\Users\UserName\AppData\Local\Continuum\Anaconda2\python.exe -u -c "import setuptools, tokenize;__file__='c:\\users\\UserName\\appdata\\local\\temp\\pip-build-gwroh2\\cupy\\setup.py';
f=getattr(tokenize, 'open', open)(__file__);code=f.read().replace('\r\n', '\n');f.close();
exec(compile(code, __file__, 'exec'))" install --record c:\users\UserName\appdata\local\temp\pip-tbtixt-record\install-record.txt --single-version-externally-managed --compile" failed with error code 1 in c:\users\UserName\appdata\local\temp\pip-build-gwroh2\cupy\

Failed to buil cupy I checked nvcc version

nvcc -V

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2016 NVIDIA Corporation
Built on Mon_Jan__9_17:32:33_CST_2017
Cuda compilation tools, release 8.0, V8.0.60

The same problem occurs when I try installing cupy by pip with chainer1.22 environment

This is my environment

Windows 10 pro 64bit
Anaconda 2.4.11
NVDIA GTX 1060 6GB, Driver Version: 398.11
CUDA 8.0
Visual studio 2015

I referenced this post But it doesnt work for me.

HOW CAN I IMPORT CUPY WITH CHAINER4.1.0?

Thank you for reading my question!

score 3 · Accepted Answer · answered Jun 22 '18 at 09:11

3

As far as I know, chainer and cupy become separated after chainer version 2. So cupy is automatically installed with chainer version 1, but you need to install cupy separately after chainer version 2. So you need to install cupy but I don't know the reason why build failed.

If you can understand the CUDA version which you are using, you can install from built package cupy-cudaXX where XX represents your CUDA version. Try below:

# make sure cupy is uninstalled
pip uninstall cupy
pip uninstall cupy
# based on the cuda version, install command changes.
# Ex. CUDA version is 8.0
pip install cupy-cuda80
# Ex2. CUDA version is 9.0
# pip install cupy-cuda90

answered Jun 22 '18 at 09:11

corochann

1,604
1
13
24

Thank you for your answer. We tried to install cupy with your recommend, but it still doesn't work. Error: "Could not find a version that satisfies the requirement cupy-cuda80 (from versions: ) No matching distribution found for cupy-cuda80" – satoshi-p Jun 22 '18 at 10:01
From the error message, I guess your pip is too old to install precompiled binary. Try updating pip with `pip install -U pip`. – kmaehashi Jul 11 '18 at 04:10

score 3 · Answer 2 · answered Jul 09 '18 at 01:57

3

Try

pip install -U pip setuptools
pip freeze
pip uninstall chainer cupy cupy-cuda80 cupy-cuda90 cupy-cuda91 cupy-cuda92
pip install --no-cache-dir cupy-cuda80

answered Jul 09 '18 at 01:57

Yuki Hashimoto

1,013
7
19

you would need to install chainer along with the correct version of cupy after you uninstall everything -- pip install chainer cupy-cuda90 – Nilav Baran Ghosh Feb 11 '19 at 20:01

Khan · Answer 3 · 2018-10-11T14:11:46.583

0

i recently encountered this error on windows environment python 3.6 vs 2015 cuda 8.0. i tried

pip install cupy

but it didnt installed the package. i guess because for cupy we have to mention the cuda version due to the compatilbility issues. However, entering

pip install cupy--cuda80

installed the package properly.

edited Oct 11 '18 at 14:11

answered Oct 02 '18 at 12:13

Khan

1,288
12
11

Can not import cupy with chainer 4.1.0

3 Answers3