lru_cache on instance methods

[bpdavis86]

This repo is good work, thanks for providing a valuable resource for tracking researchers!

I noticed you are using lru_cache for instance methods in the repo, e.g. in updater.kalman.KalmanUpdater.predict_measurement. It has been noted that this can cause memory leaks due to cyclic dependency in the garbage collection (the cache never releases reference to the instance).

https://stackoverflow.com/questions/33672412/python-functools-lru-cache-with-instance-methods-release-object

In my own code, I've worked around this by implementing a plain function (or static method) that is lru_cached and then redirecting the instance method to it. Some alternative solutions are provided at the above post. (I see you've used the staticmethod technique in some places as well).

Don't know if you have actually encountered issues with this in practice, but I thought I'd bring it up.

-Ben Davis

[sdhiscocks]

Thanks for raising this @bpdavis86. It is something we were aware of, but as we don't create many Predictor or Updater objects, shouldn't really be an issue. I think looking at a per instance cache solution is probably best.