Share persistent values among processes

[muupan]

so that BatchNormalization etc can be safely shared among processes. This is required by #477

I'll remove [WIP] once I confirm that this enables training with batch normalization.

• Add chainerrl.misc.namedpersistent, a function that mimics Add namedpersistent chainer#6788
• Modify chainerrl.misc.async_.extract_params_as_shared_arrays and chainerrl.misc.async_.set_shared_params so that not only parameters but also persistent values are shared among processes.
• Add code that checks if all the extracted values are used, and fix a bug in test found by the check.

[muupan]

I confirmed that this enabled training with batch normalization and actor-learner parallelism, so I removed [WIP].

[toslunar]

def _namedchildren(link):
    ...
    elif isinstance(link, chainer.ChainList):
        for idx, child in enumerate(link._children):
            yield str(idx), child

could avoid repeating the recursion logic.

[muupan]

I'm not sure how. Can you give me some more detail? chainer/chainer#6788 is also implemented with recursion, and I don't think we need to deviate from it.

[toslunar]

Probably I meant that a mutual recursion between namedpersistent and _namedchildren looks cleaner. It can be a matter of taste. Sorry that my previous comment is not readable (to me neither at first sight).

[muupan]

Understood, thanks! I modified accordingly, so can you check again?

[muupan]

/test

[pfn-ci-bot]

Successfully created a job for commit 4aaf766:

• Dashboard for commit 4aaf766

[toslunar]

LGTM

[muupan]

/test

[pfn-ci-bot]

Successfully created a job for commit b82b710:

• Dashboard for commit b82b710