Created by: wangyang59

issue #417 (closed) Changed the usage of weights in ExpandConvBaseLayer to match the behaviour in CudnnConvLayer so that the model trained with both methods can be used interchangeably.

Added a test_ConvUnify to verify the above desired property.