Deep Classifier is an implementation of Scikeras's KerasClassifier to feed a Scikit's GridSearchCV, it is possible to compare through cross-validation several different hyperparams and layer architectures just providing them as a param_grid.
The typical architecture parameter will be a dict that contains as keys Layers, ActivationFunctions and Neurons, the values are arrays that contain the type of layer (Dense or LSTM for now), the type of activation functions (all supported by Keras) and the number of neurons (int).
A possible param_grid configuration can look like the following dictionary:
param_grid = { "optimizer__learning_rate": [0.00001, 0.1], "model__architecture": [ { "Layers": ["Dense", "Dense"], "ActivationFunctions": ["relu", "sigmoid"], "Neurons": [20, 1], }, { "Layers": ["Dense", "Dense", "Dense"], "ActivationFunctions": ["relu", "relu", "sigmoid"], "Neurons": [50, 20, 1], } ]
Here the idea is to test which one of these two architectures will perform better, also with two possible learning rates.
An example is provided with credit risk data in Classification.ipynb.
*To avoid all the prints in the console, change verbose to 0 in GridSearchCV and in DeepClassifier
-
Install requirements with
pip install -r requirements.txt -
Instantiate a base model of
Deep Classifierwith the default params that you defined, it is possible to add more params to ClassDeep Classifer, this can be done inDeepClassifier.py -
Provide a
param_gridwith different architectures and hyperparams that will be tested -
Obtain the results