yifeim commented on issue #11651: Add logistic regression tutorial URL: https://github.com/apache/incubator-mxnet/pull/11651#issuecomment-406199511 The tutorial looks awesome! A few comments: * Is it just me or everybody experiences training loss 8x larger than validation loss? I understand it may be possible due to unstable weights from backward update after every batch. But, still it is not my usual expectation. (Not to say that I have not seen similar issues before and I am still slightly puzzled by the root cause.) * The model class is linear, but the proposed network is a 3-layer overkill (which may be related to the larger training loss). If you want a fun problem, maybe consider a `xor` function class: https://medium.com/@jayeshbahire/the-xor-problem-in-neural-networks-50006411840b * Is this `mxnet.gluon.loss.LogisticLoss` equivalent to `SigmoidBCELoss`? Since you are using metrics, it may be worthwhile exploring `mx.metric.F1`. * There are some magic hyperparameters to be explained: a strong wd=0.01, an Xavier=2.34 initialization. As an elementary tutorial, I would try to simplify them unless they are part of the intended purposes.
---------------------------------------------------------------- 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
