locked
I want to do learning on a transformed version of my data, then apply the inverse transformation. Is this possible? RRS feed

  • Question

  • I have some data and I want to train a neural network to learn to predict it.

    The neural network seems to be able to learn better using the logarithm of the variable I am trying to predict, so I want to log the target variable, learn on that, and then exponentiate. 

    However Azure ML is not cooperating - it will not let me connect a "IleanerDotNet" to a math module. 

    Is this just a limitation of Azure? It seems like a pretty serious one if it is. 

    Wednesday, November 11, 2015 2:21 PM

Answers

  • Math module take a DataSet ... and that make sense..., what are you using to do your transformation.

    If I was you I would use an Execute R module ( or Execute Python Module) to do your transformation  using a mutate for example and then learn from that and of course you can still attache another Execute R module at the outcome of a Score model for example to reftransform your data to it original status...As well

    the output of R module can be the transformed dataset so you can still reuse your data..

    • Proposed as answer by neerajkh_MSFT Thursday, November 12, 2015 2:35 AM
    • Marked as answer by Hai Ning Tuesday, November 17, 2015 4:02 PM
    Wednesday, November 11, 2015 3:58 PM
  • Thanks - the problem was you have to apply the inverse function after the output of the "score model" module, not after "train model"

    Thanks!

    • Marked as answer by neerajkh_MSFT Friday, November 13, 2015 5:27 PM
    Wednesday, November 11, 2015 5:31 PM

All replies

  • Math module take a DataSet ... and that make sense..., what are you using to do your transformation.

    If I was you I would use an Execute R module ( or Execute Python Module) to do your transformation  using a mutate for example and then learn from that and of course you can still attache another Execute R module at the outcome of a Score model for example to reftransform your data to it original status...As well

    the output of R module can be the transformed dataset so you can still reuse your data..

    • Proposed as answer by neerajkh_MSFT Thursday, November 12, 2015 2:35 AM
    • Marked as answer by Hai Ning Tuesday, November 17, 2015 4:02 PM
    Wednesday, November 11, 2015 3:58 PM
  • Thanks - the problem was you have to apply the inverse function after the output of the "score model" module, not after "train model"

    Thanks!

    • Marked as answer by neerajkh_MSFT Friday, November 13, 2015 5:27 PM
    Wednesday, November 11, 2015 5:31 PM
  • If we have to apply the inverse function after the output of the "score model" module, does that mean we have to apply the function to the test dataset as well? (since both test and train datasets go to the "Score Model" and need to be in the same scale.)

    Monday, March 2, 2020 1:47 PM