Remove dependency from tensorflow

[Martin-Laclaustra]

Tensorflow is only required in this project for defining the hyperparameters structure.
In different architectures I found problems getting an environment with both pytorch and tensorflow in appropriate sync of libraries and python versions.
That is the motivation for this modification.

You might need to add some new dependencies to setup.py if these are not already required by the rest of the code or previous dependencies (I think everything is covered). These are the imports in hparam.py:
__future__ json numbers re six

I hope you can consider testing and merging this pull request.
Thank you.

[r9y9]

Sorry for the delay. I agree with dropping tensorflow dependency, but I am worried if we can further simlify hparams functionality. Just a quick thought, but for example, could we replace it with a simpler python class that supports 1) dictionary-like key/value initialization 2) class-like accessor 3) parsing from json ? Isn't it enough?

[Martin-Laclaustra]

I am worried if we can further simlify hparams functionality.
Just a quick thought, but for example, could we replace it with a simpler python class that supports

1. dictionary-like key/value initialization
2. class-like accessor
3. parsing from json ? Isn't it enough?

Probably a simpler class would be enough, but see the following thoughts:

• programming a new class will require effort, time, and good testing.
• the trimmed tensorflow "hparam.py" class is lean enough and it does not carry over "heavy" dependencies. I already did the task of detaching it from tensorflow, and the standalone class works well in my branch.
• I have seen other several projects (e.g. at NVIDIA/tacotron2) that also use only hparam from tensorflow. It may appear as "standard" for several projects, so having this as a drop-in replacement instead of diverging can be useful for reshuffling code with other projects.

In sum, I believe that it may be more convenient to use my patch than to program a full new solution.

[r9y9]

Could this be simply import hparam_tf?

[Martin-Laclaustra]

Just get the file hparam.py out of the subfolder, rename it to hparam_tf.py and import it directly as you suggest. I think it will work.
I just created the subfolder to add the readme where its origin is explained.

[r9y9]

Ah, I see. I misunderstood there's __init__py in the hparams_tf directory.

[r9y9]

Could you fix requirements in README as well? https://github.com/r9y9/deepvoice3_pytorch/tree/8cee593f3b198ef894e918f976c9a6f3041dc421#requirements Then I will merge.

[Martin-Laclaustra]

Could you fix requirements in README as well? https://github.com/r9y9/deepvoice3_pytorch/tree/8cee593f3b198ef894e918f976c9a6f3041dc421#requirements Then I will merge.

Done!

[r9y9]

Thank you for your contribution!