You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This refers to the prerelease of 0.12 from Git, commit b85978e
The example in the minimizeAutoencoder documentation doesn't run as coded. It's missing a comma, and after that fix, it gives errors about non-conformable arrays.
This is the closest version I was able to get to run:
However this produces MSE which never decreases, and warnings about NaNs, so I think I'm still doing it wrong. Could you provide a working example of how to use minimizeAutoencoder properly?
The text was updated successfully, but these errors were encountered:
thanks for your feedback. There seems to be a bug in the minimizeAutoencoder function. It is planned to remove both minimizeAutoencoder and minimizeClassifier and use the optim function in R instead to provide CG support.
Until then, it is unlikely that this issue will be completely resolved since I don't have the time nor the knowledge to fix it and I'm not the original author of the CG code (any comments on what could be going wrong here are appreciated). However, I will go through the history of the minimizeAutoencoder function and see if this bug was introduced by me or was always present.
Additionally, I will
improve the example of the minimizeAutoencoder you mentioned, it has several issues apart from the missing comma
"fix" the problem with the NaNs you encountered (minimizeAutoencoder uses the log function to determine the error, which cannot deal with negative values; I write "fix" because I simply surround the values passed to log with a call to abs, which most likely is problematic in its own way)
Sorry that I can't really help you here! :/
As a side note, autoencoders can also be trained as regression problems using backpropagation.
This refers to the prerelease of 0.12 from Git, commit b85978e
tmp<-darch(iris[,1:4], iris[,1:4], c(4,10,2,10,4), darch.isClass=F, darch.fineTuneFunction = "minimizeAutoencoder", darch.numEpochs=20,darch.unitFunction=tanhUnit,preProc.params = list(method=c('center','scale')))
However this produces MSE which never decreases, and warnings about NaNs, so I think I'm still doing it wrong. Could you provide a working example of how to use minimizeAutoencoder properly?
The text was updated successfully, but these errors were encountered: