Mike Dusenberry
Mike Dusenberry
Also reproducibility tests if we can do it with TF.
TFP has been updated such that `tfp.distributions.*` objects can be initialized in a tape-safe manner. I.e., it's now possible to create a distribution within one tape, and use it in...
Currently, ed RandomVariable sampling for reparameterized distributions can encounter errors in mixed-precision settings (such as bfloat16) due to a mismatch between the the dtype of output of the `sample`computation in...
This is a work-in-progress for a set of functions to convert our models to fully-convolutional versions in order to be able to make predictions on the full images for increased...