[BUG] cross_entropy_with_selfnorm
Created by: yu239-zz
Currently this cost layer assumes that the input is already softmax normalized (i.e., sum-to-one). Thus the second line of the following code is wrong:
Matrix::resizeOrCreate(sftMaxSum_, output.getHeight(), 1, false, useGpu_);
output.rowSum(*sftMaxSum_);
sftMaxSum_->log2();
Because the rowSum is always 1. I think this layer needs the non-normalized vector as the input so that it is able to sum the normalizer.