The evaluation is going to be made on `mean_squared_error`, keep that in mind while training.