unrealwill commented on issue #8025: How to use softlabel in MXNET?
URL: 
https://github.com/apache/incubator-mxnet/issues/8025#issuecomment-332154231
 
 
   @xuliuac 
   I'm not quite familiar with mxnet yet, but I think that metric and losses 
are two different things. If mxnet is anything like Keras, Metrics are being 
used for reporting only (additional info), and loss being used for actual 
training.
   
   (If I didn't make a sign error) The loss won't converge to 0 (you can 
calculate the optimum value the loss could attain : (sum(-log(max(label)) )) (a 
soft label is similar to training set error (two training examples with 
different label) like saying here is this example I don't know if it's a 3 or a 
8 but I'm 65% confident it's a 8 ), but it should converge towards a network 
trying to reach certainty in its prediction (a deterministic policy) (It will 
predict a 8 100%, but not get penalized too much if he predict a 3 ).
   
   If you try to learn a stochastic policy then use a Kullback-Leibler 
divergence between you prediction and your soft label.
   
   
   
   
   
 
----------------------------------------------------------------
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

Reply via email to