Add PyTorch support

[utsavgarg]

PyTorch is a recent deep learning framework which has become quite popular for researchers and we want to add support for it in Fabrik. The backend for PyTorch should be independent of our other backends so that changes in one do not affect the other. Adding support for this would entail adding both import functionality (which converts the PyTorch file to our intermediate representation) and the export functionality (which does the conversion the other way around).

Deliverables

A basic setup that fulfils the following requirements:

• [ ] Investigating whether it is better to use the existing Keras backend to handle the PyTorch import and export, or to have a separate backend for PyTorch
• [ ] Adding support for PyTorch import and export for at least as many layers as are present in the Keras backend
• [ ] Write unittests for the PyTorch backend

Please comment for if you have questions.

[PalashTanejaPro]

@utsavgarg Import using Keras is implemented in #284, can the person working on this test and use that code, if they are implementing import/export using Keras?

[utsavgarg]

@PalashTanejaPro Sure.

[thatbrguy]

@utsavgarg Can I test the feasibility of using Keras's import for PyTorch implemented in #284 ?

[utsavgarg]

@Sayan98 Sure, go ahead.

[vasu-dev]

@utsavgarg @PalashTanejaPro @thatbrguy I would like to work on this , what all features has been already assigned ??

[thatbrguy]

@vasu-dev I'm waiting until the remote pytorch branch is updated with the master, after that i'll open a PR to add the basic folders and files where everyone can contribute. I'm mostly working on the import, it's better if you ask more details on the gitter channel.

[MuLx10]

@utsavgarg @thatbrguy @PalashTanejaPro Has there been any progress on this issue? I want to work on this issue.

[mselimata]

I also want to work on this issue on GSOC

[dwlsalmeida]

How is progress on this, especially with the release of PyTorch 1.0? Is there room to help still?