yimyom commented on issue #5326: Train-accuracy=0.000000 URL: https://github.com/apache/incubator-mxnet/issues/5326#issuecomment-338124609 I'm building an autoencoder. I found my bug in fact. I was using an "accuracy" loss function instead of a "mse" (or "mae" or "rmse"). For those making the same mistake as me, "accuracy" is used to test for classification and you want the output of the network to be exactly the label of your classe. Obviously in a regression problem, the result will never be exactly the same as the observed value. So you want to use mse (or rmse or mae) instead because they compute a squared-error (root squared-error, absolute error). @edmBernard : thanks for the advice. I will indeed add a phase of layer-wise pretraining. It seems it will improve the model by a lot.
---------------------------------------------------------------- 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: us...@infra.apache.org With regards, Apache Git Services
