New modules in PyTorch

Since the area of Deep Learning is nascent, hundreds of brilliant research papers are being published every week. This requires that a Deep Learning framework should have a policy towards adding modules for the same.

  1. Is there a policy for implementing new modules in PyTorch (pytorch.nn)?
  2. Which modules do you like to add/ wish being added in PyTorch in nearby future? Please comment.
  3. If I want to implement a new module (like Batch Renormalization), what are the criteria that my implementation should fulfill in order to get incorporated in PyTroch? How should I open an issue regarding the same and submit a Pull Request? (I am relatively new to Open Source development)

In general, we promote users to build new modules in their own extensions, i.e. instead of torch.nn.BatchRenormalization, foo.BatchRenormalization. Here’s an example: https://github.com/fxia22/stn.pytorch

We will take the most popular / most used modules and carefully integrate them into the core torch.nn.
For examples of how to contribute, you can look at Pull Requests that have already been made to the pytorch core that have contributed an additional module.

4 Likes