Loss, metrics, and scoring in Keras

Question

What is the difference between loss, metrics and scoring in building a keras model? Should they be different or same? In a typical model, we use all of the three forGridSearchCV.

Here is the snapshot of a typical model for regression which uses all the three.

def create_model():

 model = Sequential()
 model.add(Dense(12, input_dim=1587, activation='relu'))
 model.add(Dense(1, activation='sigmoid'))

 model.compile(loss='mean_squared_error', optimizer='adam', metrics=['mean_squared_error'])
 return model

model = KerasRegressor(build_fn=create_model, verbose=0)
batch_size = [10, 20, 40, 60, 80, 100]
epochs = [10, 50, 100]
param_grid = dict(batch_size=batch_size, epochs=epochs)
grid = GridSearchCV(estimator=model,param_grid=param_grid, scoring='r2' n_jobs=-1)
grid_result = grid.fit(X, Y)

Answer 1

No, they are all different things used for different purposes in your code.

There are two parts in your code.

1) Keras part:

 model.compile(loss='mean_squared_error', 
               optimizer='adam', 
               metrics=['mean_squared_error'])

a) loss: In the Compilation section of the documentation here, you can see that:

A loss function is the objective that the model will try to minimize.

So this is actually used together with the optimizer to actually train the model

b) metrics: According to the documentation:

A metric function is similar to a loss function, except that the results from evaluating a metric are not used when training the model.

This is only used to report the metrics so that the used (you) can judge the performance of model. It does not impact how the model is trained.

2) Grid-search part:

scoring: Again, check the documentation

A single string or a callable to evaluate the predictions on the test set.

This is used to find the combination of parameters which you defined in param_grid which gives the best score.

They can very well (in most cases) be different, depending on what you want.