Details
-
Bug
-
Status: Resolved
-
Major
-
Resolution: Fixed
-
None
-
None
Description
The Param::HandlePutMsg doesn't set the weight decay multiplier and learning rate multiplier correctly. The current code doesn't use the values from the received messages.
This bug is found when testing the CIFAR10 example. That example uses a large weight decay for top layers to avoid over-fitting. The test is fine if parameters are updated at worker side but gets over-fitting when parameters are updated at server side. The reason is that the weight decay multiplier is not set properly.