smash.Model.nn_parameters#

property Model.nn_parameters[source]#

The weight and bias of the parameterization neural network.

The neural network is used in hybrid model structures to correct internal fluxes.

Returns:

nn_parametersNN_ParametersDT: It returns a Fortran derived type containing the weight and bias of the parameterization neural network.

See also

Model.get_nn_parameters_weight: Get the weight of the parameterization neural network.
Model.get_nn_parameters_bias: Get the bias of the parameterization neural network.
Model.set_nn_parameters_weight: Set the values of the weight in the parameterization neural network.
Model.set_nn_parameters_bias: Set the values of the bias in the parameterization neural network.

Examples

>>> from smash.factory import load_dataset
>>> setup, mesh = load_dataset("cance")

Set the hydrological module to 'gr4_mlp' (hybrid hydrological model with multilayer perceptron)

>>> setup["hydrological_module"] = "gr4_mlp"
>>> model = smash.Model(setup, mesh)

By default, the weight and bias of the parameterization neural network are set to zero. Access to their values with the getter method get_nn_parameters_weight or get_nn_parameters_bias

>>> model.get_nn_parameters_bias()
[array([0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
        dtype=float32), array([0., 0., 0., 0.], dtype=float32)]

The output contains a list of weight or bias values for trainable layers.

Set random values with the setter methods set_nn_parameters_weight or set_nn_parameters_bias using available initializers

>>> model.set_nn_parameters_bias(initializer="uniform", random_state=0)
>>> model.get_nn_parameters_bias()
[array([ 0.09762701,  0.43037874,  0.20552675,  0.08976637, -0.1526904 ,
        0.29178822, -0.12482557,  0.78354603,  0.92732555, -0.23311697,
        0.5834501 ,  0.05778984,  0.13608912,  0.85119325, -0.85792786,
        -0.8257414 ], dtype=float32),
array([-0.9595632 ,  0.6652397 ,  0.5563135 ,  0.74002427], dtype=float32)]

If you are using IPython, tab completion allows you to visualize all the attributes and methods

>>> model.nn_parameters.<TAB>
model.nn_parameters.bias_1                  model.nn_parameters.from_handle(
model.nn_parameters.bias_2                  model.nn_parameters.weight_1
model.nn_parameters.bias_3                  model.nn_parameters.weight_2
model.nn_parameters.copy()                  model.nn_parameters.weight_3

Note

Not all layer weights and biases are used in the neural network. The default network only uses 2 layers, which means that weight_3 and bias_3 are not used and are empty arrays in this case

>>> model.nn_parameters.weight_3.size, model.nn_parameters.bias_3.size
(0, 0)

To set another neural network structure

>>> setup["hidden_neuron"] = (32, 16)
>>> model_2 = smash.Model(setup, mesh)

In this case, the number of layers is 3 instead of 2

>>> weights = model_2.get_nn_parameters_weight()
>>> len(weights)
3

>>> weights[2].size
64

smash.Model.nn_parameters#

This Page