Long term saving and loading standards - using TorchScript?

[bkmi]

I propose that we operate under the assumption that all of our users will be able to instantiate the class again, then load the parameters using state_dict, as usual in pytorch. Why? This is how pytorch works and we are not better at coding that the devs of pytorch. If we try to support this, we will only answering corner cases for individuals who should be able to handle this themselves.

To address the implicit main question of using swyft in production, namely what if I want to use the ratio estimator fast and without access to the defining python code? Well there is an answer for that, TorchScript. I propose that we create a long term goal of supporting the export of our ratio estimators to TorchScript for production use, i.e., integrating into the gw detection pipeline, etc.

[bkmi]

I encountered the saving and loading standards of __getstate__ and __setstate__. Perhaps we can use those going forward?

https://realpython.com/lessons/getstate-setstate/