# Odessa Version
    # Modified version of HTW approach
    def _steering_onpu(self):
        model = Sequential()
        model.add(BatchNormalization(input_shape=(96,128,3)))

        # Convolution no.1
        model.add(
            Conv2D(
                filters=16,
                kernel_size=(5, 5),
                strides=(2, 2),
                bias_initializer='he_normal',
                padding="valid"))
        model.add(ELU())

        model.add(MaxPooling2D(pool_size=(2, 2)))
        model.add(BatchNormalization())

        # Convolution no.2
        model.add(
            Conv2D(
                filters=24,
                kernel_size=(5, 5),
                strides=(2, 2),
                bias_initializer='he_normal',
                padding="valid"))
        model.add(ELU())
        model.add(BatchNormalization())

        # Convolution no.3
        model.add(
            Conv2D(
                filters=32,
                kernel_size=(3, 3),
                bias_initializer='he_normal',
                padding="valid"))
        model.add(ELU())
        model.add(BatchNormalization())

        model.add(Flatten())

        # Fully connected no.1
        model.add(
            Dense(
                units=300,
                bias_initializer='he_normal'))
        model.add(ELU())
        model.add(BatchNormalization())
        model.add(Dropout(0.25))

        # Fully connected no.2
        model.add(
            Dense(
                units=100,
                bias_initializer='he_normal'))
        model.add(ELU())
        model.add(BatchNormalization())
        model.add(Dropout(0.25))

        # Fully connected no.3
        model.add(
            Dense(
                units=1,
                bias_initializer='he_normal',
                activation='tanh',
                name='prediction'))

        opt = optimizers.Nadam()
        model.compile(loss='mse', optimizer=opt, metrics=['mse'])
        model.summary()
        return model