Extend with MLPClassifer

[Wilsonlou]

Dear experts,

I tried to extend MLP from sklearn as a candidate model. However, it seems very hard to set search space for hidden_layer_sizes as a categorical hyper, for example:
hidden_layer_sizes = CategoricalHyperparameter(
"hidden_layer_sizes", [(20,10),(10,10)]
)

I tried OrdinalHyperparameter, but seems auto-sklearn still not supports it.

Appreciate if you can provide any help on it.

Thx

Wilson

[mfeurer]

@ahn1340 were you able to progress on the extending Auto-sklearn example?

[Wilsonlou]

Thx guys, actually the issue here is the hyperparam is a list of tuples, for other params, it's ok to work with.

[Wilsonlou]

Btw, I tried params like: Shuffle, early_stopping, nesterovs_momentum, these params are true or false. Once I tried to set like:
nesterovs_momentum = CategoricalHyperparameter(
'nesterovs_momentum', [0, 1], default_value=1
)
early_stopping = CategoricalHyperparameter(
'early_stopping', [0,1], default_value=0
)

It will show value error due to value type check in sklearn MLP code:
def _validate_hyperparameters(self):
if not isinstance(self.shuffle, bool):
raise ValueError("shuffle must be either True or False, got %s." %
self.shuffle)
https://github.com/scikit-learn/scikit-learn/blob/master/sklearn/neural_network/multilayer_perceptron.py

[mfeurer]

these params are true or false

Check the random forest hyperparameter definition and casting to learn how to use boolean hyperparameters: https://github.com/automl/auto-sklearn/blob/master/autosklearn/pipeline/components/classification/random_forest.py

actually the issue here is the hyperparam is a list of tuples

I would suggest you use integers here. One integer for the depth of the network, and one integer for the number of neurons per layer which is only active if the network is deep enough. Please check XGBoost to learn how to use so-called conditional hyperparameters: https://github.com/automl/auto-sklearn/blob/master/autosklearn/pipeline/components/classification/xgradient_boosting.py There's also a greater than condition which you can use for the neural network.

[Wilsonlou]

For hyperparam of tuples, have tried your suggestion, it works, thx:)

[mfeurer]

Closing this due to inactivity - also, there's now an example on this: http://automl.github.io/auto-sklearn/master/examples/example_extending_classification.html#sphx-glr-examples-example-extending-classification-py.