allow sharing for tensorflow and theano as numpy backends

[jcmgray]

Description

Allows sharing to be used with tensorflow and theano as numpy backends. Closes #50. The to_{backend} functions are all called internally within their respective {backend.py} submodules so I didn't try and factorize out a new to_backend function. I think it will be pretty easy (and rare) to add new conversion backends anyway. Also the tests are in test_backends.py rather than test_sharing.py just because the tensorflow tests require a session, and the theano tests can't use einsum itself.

Status

• Ready to go

[jcmgray]

Ah this has crept in as a small docs change. Can remove if necessary @dgasmith.

[codecov-io]

Codecov Report

Merging #58 into master will increase coverage by 0.04%.
The diff coverage is 100%.

[dgasmith]

LGTM. I think the doc change is a good one and we have mathjax setup already.

@fritzo Any comments?

[fritzo]

The code looks good. Do you want to add tensorflow and theano to the sharing test?

[jcmgray]

Yes that would have been nice but I've ended up adding them to test_backends.py tests for two reasons:

1. tensorflow requires a Session object to be setup and torn down between runs (which is defined with a special config in test_backends.py.
2. theano doesn't provide einsum and so can't actually perform many of the test contractions in the test_sharing.py

I didn't think too hard about unifying the tests but it is probably possible.

[fritzo]

Sounds reasonable, I think the tests are fine separated. Mostly the test_sharing.py tests ensure that the cache mechanism actually works, so it sounds reasonable to combine (1) heavy tests for a few backends + (2) smoke tests for the rest.