CODEWITHSUNDEEP
pythonscikit-learnneural-networkperceptron
Chris Parry
Chris Parry

Reputation: 3057

Early stopping and the sklearn neural_network.MLPClassifier

I am using sklearn.neural_network.MLPClassifier. I am using the early_stopping feature, which evaluates performance for each iteration using a validation split (10% of the training data by default).

However, my problem is multi-label. According to the API, validation uses subset accuracy, which is very harsh for multilabel problems.

Is it possible to define an alternative scoring function (ideally mlogloss) to be used in validation? Thanks.

Upvotes: 1

Views: 3454

Answers (1)

Chris Parry
Chris Parry

Reputation: 3057

The solution is to use sknn rather than the sklearn implementation of MLP. That allows use to add our own valid_set and specify the loss function. Details here.

Upvotes: 1

Related Questions