Created by: pengwangucla

Under paddle/gserver/layers/TransLayer.cpp

For back propagation the gradient is directly assigned:

MatrixPtr preGrad = getInputGrad(0); outputGrad->transpose(preGrad, false);

Will there be a problem when you have this layer connect to multiple branches? During the backward process, the gradient will be overwritten by the last branch, while missing the gradient from others.

Thanks.