Automatminer with using neural network

Matminer
1

When customizing pipes in automatminer, the example I see is using XGBClassifier. I would like to ask if there is a way that I can switch to a neural network for data training and how to implement it concretely.Thanks

Here are some examples from the site:
from xgboost import XGBRegressor, XGBClassifier

from automatminer import AutoFeaturizer, FeatureReducer, DataCleaner,
SinglePipelineAdaptor

autofeaturizer = AutoFeaturizer(from_preset=“production”,
cache_src="./features.json",
exclude=[“EwaldEnergy”])
cleaner = DataCleaner(max_na_frac=0.05)
reducer = FeatureReducer(reducers=(“corr”,))
learner = SinglePipelineAdaptor(classifier=XGBClassifier(n_estimators=500),
regressor=XGBRegressor(n_estimators=500))

Make a matpipe

pipe = MatPipe(
autofeaturizer=autofeaturizer,
cleaner=cleaner,
reducer=reducer,
learner=learner
)

2

All you need to do is specify two objects: a regressor and a classifier. Both of these must contain fit/predict syntax (as in sklearn BaseEstimator). Then you just put them in place of XGBClassifier(…) and XGBRegressor(…) in this example, if you’re just looking to use a single model.

Your underlying model can be anything, as long as it fits these specifications. For example, you can use a Keras Sequential model, sklearn MLP, etc.

3

The following is the simple network built by Keras. I still encountered difficulties when placing it in automatminer. I would like to ask how to operate it. I apologize for my lack of proficiency in the application of the model

QQ截图20200403095016
QQ截图202004030950161125×647 170 KB

4

Ok so generally for the SinglePipelineAdaptor you need a classifier and regressor. But if you are sure you will only be calling it with the regressor, for example, you can just use a dummy for the classifier and use your NN for the regressor.

So you would compile your model, then do:

from sklearn.dummy import DummyClassifier

nn_adapter = SinglePipelineAdaptor(regressor=model, classifier=DummyClassifier())

Or similar. I’ve just noticed unfortunately there’s no way for you to enter the extra kwargs in the pipeline (e.g., epochs=300 ), but I have opened an issue for this on github and will work on it (though it is not priority)! If you are interested in accelerating this fix, please consider contributing code!!

5

Thanks for your comments, which allow me to slove the problem. For the topic you mentioned, I think it seems to work with KerasRegressor(). Once I’ve made progress in this area, I’ll share it