Splitting the data and train the ANN
====================================
We can start by loading data and split into training set and evaluating set::
    
    def split_data():
        inputs, targets, timestamps = utils.read_input()
        
        return preprocessing.split(
            inputs,
            targets,
            timestamps,
            TRAINING_SIZE,
            EVALUATION_SIZE, 
        )


Then we only need to create ``tensorflow.Dataset`` from the splitted data and create an ANN::

    def run():
        (train_inputs, train_targets, train_timestamps), (test_inputs, test_targets, test_timestamps) = split_data()
        
        # creating Tensorflow's Dataset
        train_dataset = tf.data.Dataset.from_tensor_slices(
            (
                train_inputs.reshape(1, *train_inputs.shape),
                train_targets.reshape(1, *train_targets.shape),
            )
         )

        test_dataset = tf.data.Dataset.from_tensor_slices(
            (
                test_inputs.reshape(1, *test_inputs.shape),
                test_targets.reshape(1, *test_targets.shape),
            )
        )
    
Once the datasets are created, you can initialize the ``ModelSet`` and train it::

    # initialize a list of 25 multi-layers perceptron
    model_set = nn.ModelSet(
        nb_models=25,
        input_shape=(len(INPUTS_VAR),),
        output_shape=len(TARGETS_VAR),
    )

    # train the MLP
    model_set.fit(train_dataset, epochs=500, steps_per_epoch=5)

    # evaluate the MLP
    model_set.evaluate(test_dataset)

    # get the prediction
    prediction = np.median(model_set(test_inputs)[:, 0], axis=0)

    # save the result
    utils.write_output(pd.DataFrame(prediction, columns=TARGETS_VAR))

During the evaluation, the models that fail to estimates fluxes are deleted. The output is saved in ``my_output_dir/output/output.csv``.