Pre-processing the training data about nolearn HOT 4 CLOSED

One way to do this is to subclass NeuralNet and override the train_test_split method to use a StandardScaler in the way you describe. In this method, store the StandardScaler as an attribute on self, and access in the predict_proba method; that you'll have to override as well.

I'll be happy to hear any suggestions on making this more dynamic. In #42, I briefly discussed making train_test_split overridable with a parameter, but in your case, it seems you'd need to subclass for predict_proba anyway.

from nolearn.

Thanks for getting back to me. I think your suggestion of subclassing NeuralNet and overriding train_test_split and predict_proba is perfectly fine for my situation.

An alternative, but more involved solution, could be to add a standard_scaler=None parameter to the NeuralNet. Then, in train_test_split, one could check if self.standard_scaler is not None and then use it to fit_transform the training set X_train and then transform the validation set X_valid. The same check and transform would have to be done in predict_proba.

from nolearn.

For my use case, I decided that it would be simpler to implement it as described in my post above. Here is a link to the code in case anyone else wants to do something similar: hjweide@7f30634

Any suggestions for improvements are also welcome.

from nolearn.

The TrainSplit interface has since been added which should give you a good opportunity to apply correct scaling.

from nolearn.