chinakook commented on a change in pull request #7944: Update softmax_output.cc
URL: https://github.com/apache/incubator-mxnet/pull/7944#discussion_r141374799
##########
File path: src/operator/softmax_output.cc
##########
@@ -62,7 +62,7 @@ MXNET_REGISTER_OP_PROPERTY(SoftmaxOutput, SoftmaxOutputProp)
.. math:: \text{CE(label, output)} = - \sum_i \text{label}_i
\log(\text{output}_i)
- - The gradient of cross entropy loss w.r.t softmax output:
+ - The gradient of cross entropy loss w.r.t the input array:
.. math:: \text{gradient} = \text{output} - \text{label}
Review comment:
As I've searched, the gradient will give a correction(i.e. gradient) which
can be represented by the softmax activation to the features before softmax
output(activation), not the softmax output(activation).
https://math.stackexchange.com/questions/945871/derivative-of-softmax-loss-function
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
