related to #1

We need to test protobuf3 beta.

Numpy 1.11 is now available.

We need to make a test which tries to install chainer with various situation.

• setuptools version variation
• pip version variation
• with old cython

We need to record performance of each version.

For example, range differs in Python 2 and Python 3.

Now we have a test for snippets in the document. We need to run this test in the CI environment.
chainer/chainer#395

In version 1.4.0, chainer uses pyhdf5 and we need to install libhdf5 before install in Dockerfiles.
see chainer/chainer#459

cudnn v4 is available.
https://developer.nvidia.com/cudnn

We need to test with various environments.

• Without c compilers #132
• With anaconda #176
• When a user installed cudnn both system path and his own home directry

It uses apt-get command. Check OS and change command.

To test ImagaDataset, we need to install pillow.

A user can use Chainer on an environment without gcc if he doesn't need CUDA.
Currently we always test chainer with gcc. We need to check such a environment.

Try to install distribution file with pip command.

In example test script, data size is too small (only one data) to run mini-batch. We need to reduce mini-batch size to test it.

docker command makes files belonging to a root user. We currently chown these files after tests on the test scripts, but when a process is killed unexpectedly these files remain. It causes a problem to Jenkins because Jenkins process cannot remove files of the root user.
To prevent this problem, we need to change user before running the tests.

We now have too many Dockerfiles. We need to manage them with a script to generate Dockerfiles or use dependency.

I found a link in a comment of the repository.
NVIDIA/nvidia-docker#55 (comment)

We need to write more detailed description.

To check regression, we should test with unreleased versions of libraries and CPython. It is not feasible to test with HEADs, so we should test with beta versions or release candidates.

In cupy branch, cuda-deps is removed. So, this test script fails.

We need to use docker kill command to kill a process on docker container.

Compilation of python and cython is too slow. ccache helps it.

ref: chainer/chainer#1738

Currently we use docker command with --device option. It requires heavy configuration on Jenkins. Instead I want to use nvidia-docker and to make a script which runs nvidia-docker command.

It generates too long log messages, and fills log of jenkins.

It may be very slow as a model file is very large.

We need to run setup.py build before nosetests for Cython.

Chainer doesn't work with python 3.5.0. Python 3.5.1 is required.

CUDA 8 is now available.

Use pillow instead of opencv
chainer/chainer#158

The current test scripts uses mainly gpu=0. It causes memory allocation error when a user runs many processes even if the machine has multiple gpus.
I want to select main GPU ID based on process id.

We need to...

• change parameters of tests to reduce their size #4
• remove dependency to python-opencv #5
• add proxy to make download fast

https://github.com/chainer/

To test backward compatibility, we decided to check if a new version of chainer supports its previous version of example.

The test script sometimes doesn't stop and gets an infinite loop. We need to set timeout in the script.

Ref: chainer/chainer#1845

Add test with 3.6.0

Now that the DCGAN example is merged to Chainer, we need to test it.