适合：使用 sklearn.model_selection.GridSearchCV 传递多个输入

Question

as explained in https://keras.io/models/model , Keras' fit function x can be

Numpy array of training data (if the model has a single input), or list of Numpy arrays (if the model has multiple inputs). If input layers in the model are named, you can also pass a dictionary mapping input names to Numpy arrays.

but it doesn't seem to be possible with sklearn.model_selection.GridSearchCV fit . Is that true? Is there any workaround?

Thanks!

Answer 1

If you want to just use grid search as in sci-kit learn, you can use the KerasClassifier wrapper. You just create your model in your function.

def create_model(dropout_rate):
    #...
model = KerasClassifier(build_fn=create_model)

param_grid = dict(dropout_rate=[0.2,0.3,0.5])
grid = GridSearchCV(estimator=model, param_grid=param_grid, n_jobs=-1)
grid_result = grid.fit(X, Y)

https://machinelearningmastery.com/grid-search-hyperparameters-deep-learning-models-python-keras/

Unfortunately, this approach is kind of limited, as you don´t have a validation after each epoch. To solve this, you need to implement your own KerasClassifier.

https://github.com/keras-team/keras/issues/4278

Answer 2

The short answers are yes and yes.

GridSearchCV checks that X and y have the same length, I assume this is required for splitting the data to do cross-validation. If you have multiple inputs and the first dimension of X represents this number of inputs this check wil fail.

There is an issue to support multiple input layers: https://github.com/keras-team/keras/issues/9001 . Also check this issue for a possible workaround.