anirban-mukherjee
commented
2 years ago I have been using TF Probability 0.15.0 module with (the current) R tfprobability 0.12 library and so far have not faced any issues or detected any errors. Hence, am closing this issue on Github.
It would be nice for others to beta test also if possible (I think many of you may be on 0.15.0 anyways). If we don't see any issues and as there are no major API changes and deprecations between 0.12.0 and 0.15.0 then perhaps we can roll forward the tfprobability version to 0.15 to indicate that the R library is compatible with TensorFlow 0.15.0 and TF 2.7.0. The reason being that a current TensorFlow Probability module, but not necessarily an older one, is compatible with current TensorFlow. Hence, it would be great if possible if tfprobability is not on the critical path for updating TensorFlow as far as possible.
We may not want to do this if our goal is to ensure that R tfprobability is 100% aligned with Python TensorFlow Probability. In this case, any new features/functions since 0.12.0 will need new R features/functions as well. But IMO, this is not crucial since there is a reference page that lists the functions that are in the R library with the obvious implication that the R feature coverage is not always 100% but that what ever is on the reference webpage works with (as far as possible) the current TensorFlow Probability module.
TensorFlow probability is now at 0.15.0, which is stable against TF 2.7.0. I am unsure about the extent of difference between 0.12.0 and 0.15.0 but if looks like there are few if any deprecations in the last few updates so if possible it would be great to be able to install and use the new version if that is feasible. Anyone with any experience using the current tfprobability library and the Python 0.15.0 module? We may not have the R accessor functions for all new features but at least we can benefit from the improvements to stability and speed in the new versions.
Anirban