roads / spond Goto Github PK

In TensorFlow, fit functionality is implemented as a method of a Model class. Should we follow the same pattern with our PyTorch models?

The chosen solution is intertwined with how we will incorporate other features such as callbacks and how we will encapsulate datasets.

Commits should be checked for appropriate style.

In general, this package should abstract out device placement as much as possible, leaving it completely in control of the API user. What is the best strategy to achieve this?

Need a set of standardized benchmarks to facilitate comparison of different alignment algorithms.

1. A "sanity check" benchmark that involves two 2D systems with 10 concepts that are exactly the same.
2. A medium difficulty benchmark with 2D systems with 30 concepts with both a rotation and some random noise in order to map from one to the other.
3. A difficult benchmark.

It would be very convenient if we had a short summaries of all the papers that have guided and informed our research. This would be useful for new and existing team members. Existing team members have a place to look when they are trying to remember what paper did what. New team members have a guide to the literature to help bring them up to speed.

I think it would be reasonable to place this content in docs. Eventually this could be part of the "readthedocs" page.

Need a standardized way to keep track of all the various models, tasks, and results.

Designing an arbitrary N-system alignment algorithm is tricky. What are some good abstractions for this class of alignment algorithms?