Just need to implement two new ops: randn and rademacher, then use them to implement flipout in a backend-independent way in Dense. Can implement rademacher for pytorch via 2*torch.randint(0, 1, size)-1.
Also, in the TF implementation, update tfp.python.math.random_rademacher (which is deprecated) to tfp.random.rademacher.
Just need to implement two new ops:
randn
andrademacher
, then use them to implement flipout in a backend-independent way in Dense. Can implement rademacher for pytorch via2*torch.randint(0, 1, size)-1
.Also, in the TF implementation, update
tfp.python.math.random_rademacher
(which is deprecated) totfp.random.rademacher
.