google/edward2

Add Sampling option to Dropout methods in edward2/baseline

jereliu opened this issue · 1 comments

Currently the dropout method in edward2/baseline generates only 1 sample during prediction, which can be the reason behind the fact that it did not as good performance as the original paper. It is worth investigating whether sampling from dropout posterior can improve the performance.

A possibly related GitHub Issue: Unable to reproduce the accuracy of WRN-28-10 on Cifar-100

Implemented in commit 990e3e7