fix shape issues, improve performance, make it pip installable, etc

[carusyte]

This PR is intended to be (hopefully) a better organized one to #35 ...
Recap:

1. Pip installable: One can generate installable packages by running python setup.py sdist under the root directory of this repo and then install it by pip install ./dist/dnc-0.0.2.tar.gz. Some versioning and author information may not be correct, which can be edited or enriched in the setup.py file.
2. tf.unstack can't process tensors with partially-known shape info. Methods in the util.py are refactored by utilizing a combination of tensor operations as a workaround, meanwhile ensuring those operations can be run in GPU.
3. Performance. Some operations (or its gradients) can't be run in GPU as of tensorflow version 1.9(e.g. tf.reduce_prod, see #3957,#8841, and tf.nn.top_k, see #5719).
4. Replace deprecated APIs. Some methods/function arguments are deprecated as of tf v1.9

[dm-jrae]

Thanks so much for addressing the issues - it's looking great, no worries re. the pull request.

Just want to check on the train script though, it looks like there's no dnc import right now and it should fail. Can you just double check this (haven't unit tested the training script).

[dm-jrae]

I think you want this import in, no?

[carusyte]

Sorry I just missed this one. I'm short of budgets and all machines(only 2 actually) at my disposal are busy running something, so I didn't bother to run the script...

Now it seems able to start and run smoothly, but I don't have the computer power to run it through.

If there's any other issues, please don't hesitate to point them out.

[dm-jrae]

Ok thanks for the fix :-)