Train multiple neural networks in one Analysis?

Thomas_K

The title basically says it all. I want to try different hyperparameters for my Neural Network (or algorithms in general). For some, like random forest, I can specify a list - e.g., max_depth. What I need is a queue of Neural Networks with different hyperparameters, so that I can start them in the evening and come back to the results in the morning.

How to do this?

Alex_Combessie

Hello,

It is not possible at the moment on the visual interface.

Instead, for hyperparameter optimization on neural networks, we invite you to code your own custom Python model (in the Analysis > Design > Algorithms section). For instance, for a neural network from scikit-learn (MLP), you can use this:

from sklearn.neural_network import MLPClassifier
from sklearn.model_selection import GridSearchCV

parameters={
'alpha': [1,10,0.1],
'activation': ["logistic", "relu"]
}

mlp = MLPClassifier()

clf = GridSearchCV(
    estimator=mlp,
    param_grid=parameters,
    n_jobs=-1,
    verbose=2,
    cv=5
    )

Note that we are looking to integrate neural networks more deeply into our product. We will keep you posted!

Cheers,

Alex

Thomas_K

Thanks! I will try the custom version as soon as the current NN classifier is finished.

Thomas_K

Unfortunately, I get the following error:
NameError: name 'gridSearchCV' is not defined
--> Fixed by using GridSearchCV instead.

After fixing this, I now get:
get_params() must be called with MLPClassifier instance as first argument.
and I'm not really sure what to do about that.

Alex_Combessie

Yes, you first need to declare an instance of MLPClassifier: `mlp=MLPClassifier()` before it can be passed to GridSearchCV.