Fix PR from FC first review

smash/_constant.py

@@ -589,7 +589,7 @@ def get_rr_internal_fluxes_from_structure(structure: str) -> list[str]:
 ### OPTIMIZABLE NN PARAMETERS ###
 #################################
 
-OPTIMIZABLE_NN_PARAMETERS = ["weight_1", "bias_1", "weight_2", "bias_2"]
+NN_PARAMETERS_KEYS = ["weight_1", "bias_1", "weight_2", "bias_2"]
 
 ### SETUP ###
 #############
@@ -995,7 +995,7 @@ def get_rr_internal_fluxes_from_structure(structure: str) -> list[str]:
     "atmos_data": ["mean_prcp", "mean_pet", "mean_snow", "mean_temp"],
     "rr_parameters": ["keys", "values"],
     "rr_initial_states": ["keys", "values"],
-    "nn_parameters": OPTIMIZABLE_NN_PARAMETERS,
+    "nn_parameters": NN_PARAMETERS_KEYS,
     "serr_mu_parameters": ["keys", "values"],
     "serr_sigma_parameters": ["keys", "values"],
     "response": ["q"],

smash/core/model/_build_model.py

@@ -12,7 +12,7 @@
     DEFAULT_RR_PARAMETERS,
     DEFAULT_SERR_MU_PARAMETERS,
     DEFAULT_SERR_SIGMA_PARAMETERS,
-    OPTIMIZABLE_NN_PARAMETERS,
+    NN_PARAMETERS_KEYS,
     SERR_MU_MAPPING_PARAMETERS,
     SERR_SIGMA_MAPPING_PARAMETERS,
     STRUCTURE_ADJUST_CI,
@@ -181,7 +181,7 @@ def _build_parameters(
 
     # % Initalize weights and biases of ANN if hybrid model structure is used
     if sum(setup.neurons) > 0:
-        for key in OPTIMIZABLE_NN_PARAMETERS:
+        for key in NN_PARAMETERS_KEYS:
             # zero init
             setattr(parameters.nn_parameters, key, 0)
 

smash/core/model/_standardize.py

@@ -416,7 +416,7 @@ def _standardize_model_setup_descriptor_name(descriptor_name: ListLike | None, *
 
 
 def _standardize_model_setup_hidden_neuron(hidden_neuron: Numeric, **kwargs) -> int:
-    if isinstance(hidden_neuron, (int, float, np.number)):
+    if isinstance(hidden_neuron, (int, float)):
         hidden_neuron = int(hidden_neuron)
         if hidden_neuron <= 0:
             raise ValueError("hidden_neuron model setup must be a positive number")
@@ -763,7 +763,9 @@ def _standardize_set_nn_parameters_weight_value(
                             f"from shape {arr.shape} into shape {weights[i].shape}"
                         )
                 else:
-                    raise TypeError("Each element of value argument must be float or a Numpy array")
+                    raise TypeError(
+                        "Each element of value argument must be of Numeric type (int, float) or np.ndarray"
+                    )
 
     else:
         raise TypeError("value argument must be a list of a same size with layers")
@@ -797,7 +799,9 @@ def _standardize_set_nn_parameters_bias_value(
                             f"from shape {arr.shape} into shape {biases[i].shape}"
                         )
                 else:
-                    raise TypeError("Each element of value argument must be float or a Numpy array")
+                    raise TypeError(
+                        "Each element of value argument must be of Numeric type (int, float) or np.ndarray"
+                    )
 
     else:
         raise TypeError("value argument must be a list of a same size with layers")

smash/core/model/model.py

@@ -1145,7 +1145,7 @@ def nn_parameters(self) -> NN_ParametersDT:
         >>> setup["hydrological_module"] = "gr4_mlp_alg"
         >>> model = smash.Model(setup, mesh)
 
-        By default, the weights and biases of the parameterization neural network is set to zero.
+        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 <Model.get_nn_parameters_weight>` or
         `get_nn_parameters_bias <Model.get_nn_parameters_bias>`
@@ -2394,7 +2394,7 @@ def get_nn_parameters_weight(self) -> list[NDArray[np.float32]]:
         Returns
         -------
         value : list[`numpy.ndarray`]
-            A list of arrays representing the weights for each layer of the parameterization neural network.
+            A list of arrays representing the weights of all layers.
 
         See Also
         --------
@@ -2416,7 +2416,7 @@ def get_nn_parameters_weight(self) -> list[NDArray[np.float32]]:
         >>> setup["hidden_neuron"] = 3
         >>> model = smash.Model(setup, mesh)
 
-        By default, the weights of the parameterization neural network is set to zero.
+        By default, the weights of all layers are set to zero.
         Access to their values with the getter methods
         `get_nn_parameters_weight <Model.get_nn_parameters_weight>`
 
@@ -2440,7 +2440,7 @@ def get_nn_parameters_bias(self) -> list[NDArray[np.float32]]:
         Returns
         -------
         value : list[`numpy.ndarray`]
-            A list of arrays representing the biases for each layer of the parameterization neural network.
+            A list of arrays representing the biases of all layers.
 
         See Also
         --------
@@ -2462,7 +2462,7 @@ def get_nn_parameters_bias(self) -> list[NDArray[np.float32]]:
         >>> setup["hidden_neuron"] = 6
         >>> model = smash.Model(setup, mesh)
 
-        By default, the biases of the parameterization neural network is set to zero.
+        By default, the biases of all layers are set to zero.
         Access to their values with the getter methods
         `get_nn_parameters_bias <Model.get_nn_parameters_bias>`
 
@@ -2486,8 +2486,8 @@ def set_nn_parameters_weight(
         Parameters
         ----------
         value : list[`float` or `numpy.ndarray`] or None, default None
-            The list of value(s) to set to the weights of the neural network.
-            If the value is a `numpy.ndarray`, its shape must be broadcastable into the weight shape.
+            The list of values to set to the weights of all layers. If an element of the list is
+            a `numpy.ndarray`, its shape must be broadcastable into the weight shape of that layer.
             If not used, a default or specified initialization method will be used.
 
         initializer : str, default 'glorot_uniform'
@@ -2580,8 +2580,8 @@ def set_nn_parameters_bias(
         Parameters
         ----------
         value : list[`float` or `numpy.ndarray`] or None, default None
-            The list of value(s) to set to the biases of the neural network.
-            If the value is a `numpy.ndarray`, its shape must be broadcastable into the bias shape.
+            The list of values to set to the biases of all layers. If an element of the list is
+            a `numpy.ndarray`, its shape must be broadcastable into the bias shape of that layer.
             If not used, a default or specified initialization method will be used.
 
         initializer : str, default 'zeros'

smash/core/simulation/_doc.py

@@ -1434,7 +1434,11 @@ def _gen_docstring_from_base_doc(
           ``'llr-slope-b'``, ``'ht-dd-a'``).
         - ``<key>-<code>``: Structural error parameter where ``<key>`` is the name of any structural error mu
           or sigma parameters and ``<code>``, the corresponding gauge (``'sg0-V3524010'``, ``'sg1-V3524010'``,
-          etc)
+          etc).
+        - ``<key>-<row>-<col>``: Weights and biases of the parameterization neural network where ``<key>``
+          indicates the layer and type of parameter (e.g., ``'weight_1'`` for the first layer weights,
+          ``'bias_2'`` for the second layer biases), and ``<row>``, ``<col>`` represent the corresponding
+          position in the matrix or vector (``'weight_2-23-21'``, ``'bias_1-16'``, etc).
 
     - x_bkg : `numpy.ndarray`
         An array of shape *(n,)* containing the background values of the control vector.
@@ -1605,11 +1609,12 @@ def _gen_docstring_from_base_doc(
 _smash_bayesian_optimize_doc_substitution = DocSubstitution(
     model_parameter="model : `Model`\n\tPrimary data structure of the hydrological model `smash`.",
     default_optimize_options_func="default_bayesian_optimize_options",
-    parameters_nn_parameters="",
-    parameters_note_nn_parameters="",
+    parameters_nn_parameters="- `Model.nn_parameters`, if using a hybrid structure model "
+    "(depending on **hydrological_module**)",
+    parameters_note_nn_parameters=", `Model.nn_parameters` (if used)",
     parameters_serr_mu_parameters="- `Model.serr_mu_parameters`",
     parameters_serr_sigma_parameters="- `Model.serr_sigma_parameters`",
-    parameters_note_serr_parameters=", `Model.serr_mu_parameters` and `Model.serr_sigma_parameters`",
+    parameters_note_serr_parameters=", `Model.serr_mu_parameters`, `Model.serr_sigma_parameters`",
     bounds_get_serr_parameters_bounds=", `Model.get_serr_mu_parameters_bounds` and "
     "`Model.get_serr_sigma_parameters_bounds`",
     model_return="model : `Model`\n\t It returns an updated copy of the initial Model object.",
@@ -1620,11 +1625,12 @@ def _gen_docstring_from_base_doc(
 _model_bayesian_optimize_doc_substitution = DocSubstitution(
     model_parameter="",
     default_optimize_options_func="default_bayesian_optimize_options",
-    parameters_nn_parameters="",
-    parameters_note_nn_parameters="",
+    parameters_nn_parameters="- `Model.nn_parameters`, if using a hybrid structure model "
+    "(depending on **hydrological_module**)",
+    parameters_note_nn_parameters=", `Model.nn_parameters` (if used)",
     parameters_serr_mu_parameters="- `Model.serr_mu_parameters`",
     parameters_serr_sigma_parameters="- `Model.serr_sigma_parameters`",
-    parameters_note_serr_parameters=", `Model.serr_mu_parameters` and `Model.serr_sigma_parameters`",
+    parameters_note_serr_parameters=", `Model.serr_mu_parameters`, `Model.serr_sigma_parameters`",
     bounds_get_serr_parameters_bounds=", `Model.get_serr_mu_parameters_bounds` and "
     "`Model.get_serr_sigma_parameters_bounds`",
     model_return="",
@@ -1650,11 +1656,12 @@ def _gen_docstring_from_base_doc(
 _bayesian_optimize_control_info_doc_appender = DocAppender(_bayesian_optimize_control_info_doc, indents=0)
 _smash_bayesian_optimize_control_info_doc_substitution = DocSubstitution(
     default_optimize_options_func="default_bayesian_optimize_options",
-    parameters_nn_parameters="",
-    parameters_note_nn_parameters="",
+    parameters_nn_parameters="- `Model.nn_parameters`, if using a hybrid structure model "
+    "(depending on **hydrological_module**)",
+    parameters_note_nn_parameters=", `Model.nn_parameters` (if used)",
     parameters_serr_mu_parameters="- `Model.serr_mu_parameters`",
     parameters_serr_sigma_parameters="- `Model.serr_sigma_parameters`",
-    parameters_note_serr_parameters=", `Model.serr_mu_parameters` and `Model.serr_sigma_parameters`",
+    parameters_note_serr_parameters=", `Model.serr_mu_parameters`, `Model.serr_sigma_parameters`",
     bounds_get_serr_parameters_bounds=", `Model.get_serr_mu_parameters_bounds` and "
     "`Model.get_serr_sigma_parameters_bounds`",
 )

smash/core/simulation/_standardize.py

@@ -27,7 +27,7 @@
     MAPPING,
     MAPPING_OPTIMIZER,
     METRICS,
-    OPTIMIZABLE_NN_PARAMETERS,
+    NN_PARAMETERS_KEYS,
     OPTIMIZABLE_RR_INITIAL_STATES,
     OPTIMIZABLE_RR_PARAMETERS,
     OPTIMIZABLE_SERR_MU_PARAMETERS,
@@ -146,24 +146,21 @@ def _standardize_simulation_optimize_options_parameters(
     available_parameters = available_rr_parameters + available_rr_initial_states
 
     if is_hybrid_structure:
-        available_parameters.extend(OPTIMIZABLE_NN_PARAMETERS)
+        available_parameters.extend(NN_PARAMETERS_KEYS)
 
     if is_bayesian:
         available_parameters.extend(available_serr_mu_parameters + available_serr_sigma_parameters)
 
     if parameters is None:
+        default_parameters = available_rr_parameters
+
         if is_bayesian:
-            parameters = np.array(
-                available_rr_parameters + available_serr_mu_parameters + available_serr_sigma_parameters,
-                ndmin=1,
-            )
-        elif (
-            is_hybrid_structure
-        ):  # hybrid structure is currently not working with bayes optim. TODO for later versions
-            parameters = np.array(available_rr_parameters + OPTIMIZABLE_NN_PARAMETERS, ndmin=1)
+            default_parameters += available_serr_mu_parameters + available_serr_sigma_parameters
 
-        else:
-            parameters = np.array(available_rr_parameters, ndmin=1)
+        if is_hybrid_structure:
+            default_parameters += NN_PARAMETERS_KEYS
+
+        parameters = np.array(default_parameters, ndmin=1)
 
     else:
         if isinstance(parameters, (str, list, tuple, np.ndarray)):
@@ -192,7 +189,7 @@ def _standardize_simulation_optimize_options_parameters(
 def _standardize_simulation_optimize_options_bounds(
     model: Model, parameters: np.ndarray, bounds: dict | None, **kwargs
 ) -> dict:
-    bounded_parameters = [p for p in parameters if p not in OPTIMIZABLE_NN_PARAMETERS]
+    bounded_parameters = [p for p in parameters if p not in NN_PARAMETERS_KEYS]
 
     if bounds is None:
         bounds = {}
@@ -1207,9 +1204,9 @@ def _standardize_simulation_optimize_options_finalize(
                         optimize_options["rr_initial_states_descriptor"][j, i] = 1
 
     # % nn parameters
-    optimize_options["nn_parameters"] = np.zeros(shape=len(OPTIMIZABLE_NN_PARAMETERS), dtype=np.int32)
+    optimize_options["nn_parameters"] = np.zeros(shape=len(NN_PARAMETERS_KEYS), dtype=np.int32)
 
-    for i, key in enumerate(OPTIMIZABLE_NN_PARAMETERS):
+    for i, key in enumerate(NN_PARAMETERS_KEYS):
         if key in optimize_options["parameters"]:
             optimize_options["nn_parameters"][i] = 1
 

smash/core/simulation/optimize/_standardize.py

@@ -101,12 +101,6 @@ def _standardize_bayesian_optimize_args(
 ) -> AnyTuple:
     func_name = "bayesian_optimize"
 
-    if "mlp" in model.setup.hydrological_module:
-        raise ValueError(
-            f"Bayesian optimization method is currently not working with the hybrid model structure "
-            f"'{model.setup.hydrological_module}'"
-        )
-
     # % In case model.set_rr_parameters or model.set_rr_initial_states were not used
     _standardize_simulation_parameters_feasibility(model)
 

smash/factory/net/_loss.py

@@ -4,7 +4,7 @@
 
 import numpy as np
 
-from smash._constant import OPTIMIZABLE_NN_PARAMETERS
+from smash._constant import NN_PARAMETERS_KEYS
 from smash.fcore._mw_forward import forward_run_b as wrap_forward_run_b
 from smash.fcore._mwd_parameters_manipulation import (
     control_to_parameters as wrap_control_to_parameters,
@@ -102,7 +102,7 @@ def _hcost_prime(
         grad_reg = np.reshape(grad_reg, (len(grad_reg), -1)).T
 
     # % Get the gradient of parameterization NN if used
-    grad_par = [getattr(parameters_b.nn_parameters, key).copy() for key in OPTIMIZABLE_NN_PARAMETERS]
+    grad_par = [getattr(parameters_b.nn_parameters, key).copy() for key in NN_PARAMETERS_KEYS]
 
     return grad_reg, grad_par
 

smash/factory/net/net.py

@@ -2,7 +2,7 @@
 
 from typing import TYPE_CHECKING
 
-from smash._constant import OPTIMIZABLE_NN_PARAMETERS, PY_OPTIMIZER, PY_OPTIMIZER_CLASS
+from smash._constant import NN_PARAMETERS_KEYS, PY_OPTIMIZER, PY_OPTIMIZER_CLASS
 from smash.factory.net._layers import Activation, Conv2D, Dense, Dropout, Flatten, Scale
 from smash.factory.net._loss import _hcost, _hcost_prime, _inf_norm
 
@@ -564,7 +564,7 @@ def _fit_d2p(
 
             # backpropagation and weights update
             if epo < epochs - 1:
-                for i, key in enumerate(OPTIMIZABLE_NN_PARAMETERS):
+                for i, key in enumerate(NN_PARAMETERS_KEYS):
                     if key in parameters:  # update trainable parameters of the parameterization NN if used
                         setattr(
                             instance.nn_parameters,

smash/fcore/derived_type/mwd_optimize_options.f90

@@ -99,8 +99,7 @@ subroutine Optimize_OptionsDT_initialise(this, setup)
         allocate (this%rr_parameters_descriptor(setup%nd, setup%nrrp))
         this%rr_parameters_descriptor = -99
 
-        !% contain 4 matrices/vectors: weight_1, bias_1, weight_2, bias_2
-        allocate (this%nn_parameters(4))
+        allocate (this%nn_parameters((size(setup%neurons) - 1)*2))
         this%nn_parameters = -99
 
         allocate (this%rr_initial_states(setup%nrrs))