lecrapaud 0.4.0__py3-none-any.whl

lecrapaud/search_space.py

@@ -0,0 +1,844 @@
+from typing import Optional
+
+# ML models
+from sklearn.linear_model import (
+    SGDRegressor,
+    LinearRegression,
+    SGDClassifier,
+    LogisticRegression,
+)
+from sklearn.tree import DecisionTreeRegressor, DecisionTreeClassifier
+from sklearn.neural_network import MLPRegressor, MLPClassifier
+from sklearn.svm import LinearSVR, LinearSVC
+from sklearn.naive_bayes import GaussianNB
+
+# Ensemble models
+from lightgbm import LGBMRegressor, LGBMClassifier
+from xgboost import XGBRegressor, XGBClassifier
+from sklearn.ensemble import (
+    RandomForestRegressor,
+    AdaBoostRegressor,
+    RandomForestClassifier,
+    AdaBoostClassifier,
+    BaggingClassifier,
+)
+import lightgbm as lgb
+import xgboost as xgb
+
+# DL models
+from keras import Model, Input
+from keras.layers import (
+    Dense,
+    LSTM,
+    Bidirectional,
+    GRU,
+    LayerNormalization,
+    RepeatVector,
+    MultiHeadAttention,
+    Add,
+    GlobalAveragePooling1D,
+    Dropout,
+    Activation,
+    TimeDistributed,
+)
+from tcn import TCN
+from keras.initializers import Identity
+from keras.regularizers import L2
+from keras.activations import sigmoid
+
+# Search spaces
+from ray import tune
+
+# we cannot use tune.sample_from function to make conditionnal search space, because hyperopt and bayesian opt need a fixed search space
+
+ml_models = [
+    {
+        "model_name": "linear",
+        "recurrent": False,
+        "need_scaling": True,
+        "classification": {
+            "create_model": LogisticRegression,
+            "search_params": {
+                "penalty": tune.choice(
+                    ["l2"]
+                ),  # None is not compatible with liblinear, and l1/elasticnet not compatible with most solvers
+                "C": tune.loguniform(1e-4, 1e2),
+                "l1_ratio": tune.quniform(0.2, 0.8, 0.1),
+                "solver": tune.choice(
+                    [
+                        "sag",
+                        "saga",
+                        "liblinear",
+                        "lbfgs",
+                        "newton-cg",
+                        "newton-cholesky",
+                    ]
+                ),
+                "max_iter": tune.randint(100, 1000),
+                "n_jobs": -1,
+                "random_state": 42,
+            },
+        },
+        "regression": {
+            "create_model": LinearRegression,
+            "search_params": {
+                "n_jobs": -1,
+                "fit_intercept": tune.choice([True, False]),
+            },
+        },
+    },
+    {
+        "model_name": "sgd",
+        "recurrent": False,
+        "need_scaling": True,
+        "classification": {
+            "create_model": SGDClassifier,
+            "search_params": {
+                "loss": tune.choice(
+                    [
+                        "hinge",
+                        "log_loss",
+                        "modified_huber",
+                        "squared_hinge",
+                    ]
+                ),
+                "penalty": tune.choice(["l1", "l2", "elasticnet"]),
+                "alpha": tune.loguniform(1e-6, 1e-2),
+                "l1_ratio": tune.quniform(0.2, 0.8, 0.1),
+                "max_iter": tune.randint(1000, 5000),
+                "shuffle": tune.choice([True, False]),
+                "random_state": 42,
+            },
+        },
+        "regression": {
+            "create_model": SGDRegressor,
+            "search_params": {
+                "penalty": tune.choice(["l1", "l2", "elasticnet"]),
+                "alpha": tune.loguniform(1e-6, 1e-2),
+                "l1_ratio": tune.quniform(0.2, 0.8, 0.1),
+                "max_iter": tune.randint(1000, 5000),
+                "random_state": 42,
+            },
+        },
+    },
+    {
+        "model_name": "naive_bayes",
+        "recurrent": False,
+        "need_scaling": False,
+        "classification": {
+            "create_model": GaussianNB,  # Naive Bayes classifier for classification
+            "search_params": {
+                "var_smoothing": tune.loguniform(
+                    1e-9, 1e-6
+                )  # Smoothing parameter to deal with zero probabilities
+            },
+        },
+        "regression": None,
+    },
+    {
+        "model_name": "bagging_naive_bayes",
+        "recurrent": False,
+        "need_scaling": False,
+        "classification": {
+            "create_model": BaggingClassifier,
+            "search_params": {
+                "estimator": GaussianNB(),  # Base model for bagging
+                "n_estimators": tune.randint(10, 100),  # Number of base estimators
+                "max_samples": tune.uniform(
+                    0.5, 1.0
+                ),  # Proportion of samples to draw for each base estimator
+                "max_features": tune.uniform(
+                    0.5, 1.0
+                ),  # Proportion of features to draw for each base estimator
+                "bootstrap": tune.choice(
+                    [True, False]
+                ),  # Whether samples are drawn with replacement
+                "bootstrap_features": tune.choice(
+                    [True, False]
+                ),  # Whether features are drawn with replacement
+                "random_state": 42,  # Fixed random state for reproducibility
+            },
+        },
+        "regression": None,
+    },
+    {
+        "model_name": "svm",
+        "recurrent": False,
+        "need_scaling": True,
+        "classification": {
+            "create_model": LinearSVC,
+            "search_params": {
+                # "penalty": tune.choice(["l1", "l2"]), # issue with l1 + hinge
+                "C": tune.loguniform(1e-4, 1e2),  # Regularization strength
+                "max_iter": tune.randint(100, 2000),  # Maximum number of iterations
+                "tol": tune.loguniform(1e-5, 1e-2),  # Tolerance for stopping criteria
+                "fit_intercept": tune.choice(
+                    [True, False]
+                ),  # Whether to calculate intercept
+                "loss": tune.choice(["hinge", "squared_hinge"]),  # Loss function
+                "dual": "auto",  # Dual only when hinge loss is not used and samples < features
+                "random_state": 42,  # Fixed random state for reproducibility
+            },
+        },
+        "regression": {
+            "create_model": LinearSVR,
+            "search_params": {
+                "C": tune.loguniform(1e-4, 1e2),  # Regularization strength
+                "max_iter": tune.randint(100, 2000),  # Maximum number of iterations
+                "tol": tune.loguniform(1e-5, 1e-2),  # Tolerance for stopping criteria
+                "epsilon": tune.loguniform(
+                    1e-4, 1e-1
+                ),  # Epsilon in the epsilon-insensitive loss function
+                "fit_intercept": tune.choice(
+                    [True, False]
+                ),  # Whether to calculate intercept
+                "loss": tune.choice(
+                    ["epsilon_insensitive", "squared_epsilon_insensitive"]
+                ),  # Loss function
+                "dual": "auto",  # Dual is not applicable for certain configurations in SVR
+                "random_state": 42,  # Fixed random state for reproducibility
+            },
+        },
+    },
+    {
+        "model_name": "tree",
+        "recurrent": False,
+        "need_scaling": False,
+        "classification": {
+            "create_model": DecisionTreeClassifier,
+            "search_params": {
+                "criterion": tune.choice(["gini", "entropy", "log_loss"]),
+                "max_depth": tune.randint(8, 64),
+                "min_samples_split": tune.randint(2, 10),
+                "min_samples_leaf": tune.randint(1, 4),
+                "max_features": tune.uniform(
+                    0.5, 1.0
+                ),  # Proportion of features to draw for each base estimator
+                "random_state": 42,
+            },
+        },
+        "regression": {
+            "create_model": DecisionTreeRegressor,
+            "search_params": {
+                "max_depth": tune.randint(8, 64),
+                "min_samples_split": tune.randint(2, 10),
+                "min_samples_leaf": tune.randint(1, 4),
+                "max_features": tune.uniform(
+                    0.5, 1.0
+                ),  # Proportion of features to draw for each base estimator
+                "random_state": 42,
+            },
+        },
+    },
+    {
+        "model_name": "forest",
+        "recurrent": False,
+        "need_scaling": False,
+        "classification": {
+            "create_model": RandomForestClassifier,
+            "search_params": {
+                "n_estimators": tune.randint(50, 1000),  # Number of trees in the forest
+                "max_depth": tune.randint(8, 64),  # Maximum depth of the trees
+                "min_samples_split": tune.randint(
+                    2, 20
+                ),  # Minimum samples required to split a node
+                "min_samples_leaf": tune.randint(
+                    1, 10
+                ),  # Minimum samples required at a leaf node
+                "max_features": tune.choice(
+                    ["sqrt", "log2", None]
+                ),  # Number of features to consider at each split
+                "bootstrap": tune.choice(
+                    [True, False]
+                ),  # Whether to use bootstrap sampling
+                "criterion": tune.choice(
+                    ["gini", "entropy", "log_loss"]
+                ),  # The function to measure the quality of a split
+                # "oob_score": tune.choice(
+                #     [True, False]
+                # ),  # Whether to use out-of-bag samples to estimate generalization accuracy: not working if bootstrap = False
+                "n_jobs": -1,  # Use all processors
+                "random_state": 42,  # Fixed random state for reproducibility
+            },
+        },
+        "regression": {
+            "create_model": RandomForestRegressor,
+            "search_params": {
+                "n_estimators": tune.randint(50, 1000),  # Number of trees in the forest
+                "max_depth": tune.randint(5, 30),  # Maximum depth of the trees
+                "min_samples_split": tune.randint(
+                    2, 20
+                ),  # Minimum samples required to split a node
+                "min_samples_leaf": tune.randint(
+                    1, 10
+                ),  # Minimum samples required at a leaf node
+                "max_features": tune.choice(
+                    ["sqrt", "log2", None]
+                ),  # Number of features to consider at each split
+                "bootstrap": tune.choice(
+                    [True, False]
+                ),  # Whether to use bootstrap sampling
+                "criterion": tune.choice(
+                    ["squared_error", "absolute_error", "friedman_mse"]
+                ),  # Loss function to use
+                # "oob_score": tune.choice(
+                #     [True, False]
+                # ),  # Whether to use out-of-bag samples to estimate generalization accuracy: not working if bootstrap = False
+                "n_jobs": -1,  # Use all processors
+                "random_state": 42,  # Fixed random state for reproducibility
+            },
+        },
+    },
+    {
+        "model_name": "adaboost",
+        "recurrent": False,
+        "need_scaling": False,
+        "classification": {
+            "create_model": AdaBoostClassifier,
+            "search_params": {
+                "n_estimators": tune.randint(50, 1000),  # Number of boosting stages
+                "learning_rate": tune.loguniform(
+                    1e-4, 1
+                ),  # Learning rate shrinks the contribution of each classifier
+                "random_state": 42,  # Fixed random state for reproducibility
+                "estimator": tune.choice(
+                    [
+                        DecisionTreeClassifier(max_depth=2**i, random_state=42)
+                        for i in range(1, 6)
+                    ]
+                ),  # Base estimators are decision trees with varying depths
+            },
+        },
+        "regression": {
+            "create_model": AdaBoostRegressor,
+            "search_params": {
+                "n_estimators": tune.randint(50, 1000),  # Number of boosting stages
+                "learning_rate": tune.loguniform(1e-4, 1),  # Learning rate
+                "loss": tune.choice(
+                    ["linear", "square", "exponential"]
+                ),  # Loss function to use when updating weights
+                "random_state": 42,  # Fixed random state for reproducibility
+                "estimator": tune.choice(
+                    [
+                        DecisionTreeRegressor(max_depth=2**i, random_state=42)
+                        for i in range(1, 6)
+                    ]
+                ),  # Base estimators are decision trees with varying depths
+            },
+        },
+    },
+    {
+        "model_name": "xgb",
+        "recurrent": False,
+        "need_scaling": False,
+        "classification": {
+            "create_model": "xgb",
+            "search_params": {
+                "num_boost_round": tune.randint(
+                    50, 1000
+                ),  # Number of boosting rounds (trees)
+                "early_stopping_rounds": tune.randint(5, 50),
+                "model_params": {
+                    "max_depth": tune.randint(3, 10),  # Maximum depth of trees
+                    "eta": tune.loguniform(
+                        1e-4, 0.5
+                    ),  # Learning rate, note 'eta' is used instead of 'learning_rate'
+                    "subsample": tune.quniform(
+                        0.6, 1, 0.05
+                    ),  # Subsample ratio of training instances
+                    "colsample_bytree": tune.quniform(
+                        0.6, 1, 0.05
+                    ),  # Subsample ratio of columns for each tree
+                    "gamma": tune.uniform(
+                        0, 10
+                    ),  # Minimum loss reduction for further partitioning
+                    "min_child_weight": tune.loguniform(
+                        1, 10
+                    ),  # Minimum sum of instance weights in a child
+                    "alpha": tune.loguniform(
+                        1e-5, 1
+                    ),  # L1 regularization term on weights
+                    "lambda": tune.loguniform(
+                        1e-5, 1
+                    ),  # L2 regularization term on weights
+                    "random_state": 42,  # Fixed random state
+                    "n_jobs": -1,  # Number of parallel threads for computation
+                },
+            },
+        },
+        "regression": {
+            "create_model": "xgb",
+            "search_params": {
+                "num_boost_round": tune.randint(
+                    50, 1000
+                ),  # Number of boosting rounds (trees)
+                "early_stopping_rounds": tune.randint(5, 50),
+                "model_params": {
+                    "max_depth": tune.randint(3, 10),  # Maximum depth of trees
+                    "eta": tune.loguniform(1e-4, 0.5),  # Learning rate (eta)
+                    "subsample": tune.quniform(
+                        0.6, 1, 0.05
+                    ),  # Subsample ratio of training instances
+                    "colsample_bytree": tune.quniform(
+                        0.6, 1, 0.05
+                    ),  # Subsample ratio of columns for each tree
+                    "gamma": tune.uniform(
+                        0, 10
+                    ),  # Minimum loss reduction for further partitioning
+                    "min_child_weight": tune.loguniform(
+                        1, 10
+                    ),  # Minimum sum of instance weights in a child
+                    "alpha": tune.loguniform(
+                        1e-5, 1
+                    ),  # L1 regularization term on weights
+                    "lambda": tune.loguniform(
+                        1e-5, 1
+                    ),  # L2 regularization term on weights
+                    "random_state": 42,  # Fixed random state
+                    "n_jobs": -1,  # Number of parallel threads for computation
+                },
+            },
+        },
+    },
+    {
+        "model_name": "lgb",
+        "recurrent": False,
+        "need_scaling": False,
+        "classification": {
+            "create_model": "lgb",
+            "search_params": {
+                "num_boost_round": tune.randint(
+                    50, 1000
+                ),  # Number of boosting rounds (trees)
+                "early_stopping_rounds": tune.randint(5, 50),
+                "model_params": {
+                    "max_depth": tune.randint(3, 10),  # Maximum depth of trees
+                    "learning_rate": tune.loguniform(1e-4, 0.5),  # Learning rate
+                    "num_leaves": tune.randint(
+                        20, 150
+                    ),  # Maximum number of leaves in one tree
+                    "subsample": tune.quniform(
+                        0.6, 1, 0.05
+                    ),  # Fraction of training data for each boosting round
+                    "colsample_bytree": tune.quniform(
+                        0.6, 1, 0.05
+                    ),  # Fraction of features to use per tree
+                    "min_data_in_leaf": tune.randint(
+                        20, 100
+                    ),  # Minimum number of data points in a leaf
+                    "lambda_l1": tune.loguniform(1e-5, 1),  # L1 regularization
+                    "lambda_l2": tune.loguniform(1e-5, 1),  # L2 regularization
+                    "random_state": 42,  # Fixed random state for reproducibility
+                    "n_jobs": -1,  # Use all cores for parallel computation
+                },
+            },
+        },
+        "regression": {
+            "create_model": "lgb",
+            "search_params": {
+                "num_boost_round": tune.randint(
+                    50, 1000
+                ),  # Number of boosting rounds (trees)
+                "early_stopping_rounds": tune.randint(5, 50),
+                "model_params": {
+                    "max_depth": tune.randint(3, 10),  # Maximum depth of trees
+                    "learning_rate": tune.loguniform(1e-4, 0.5),  # Learning rate
+                    "num_leaves": tune.randint(
+                        20, 150
+                    ),  # Maximum number of leaves in one tree
+                    "subsample": tune.quniform(
+                        0.6, 1, 0.05
+                    ),  # Fraction of training data for each boosting round
+                    "colsample_bytree": tune.quniform(
+                        0.6, 1, 0.05
+                    ),  # Fraction of features to use per tree
+                    "min_data_in_leaf": tune.randint(
+                        20, 100
+                    ),  # Minimum number of data points in a leaf
+                    "lambda_l1": tune.loguniform(1e-5, 1),  # L1 regularization
+                    "lambda_l2": tune.loguniform(1e-5, 1),  # L2 regularization
+                    "random_state": 42,  # Fixed random state for reproducibility
+                    "n_jobs": -1,  # Use all cores for parallel computation
+                },
+            },
+        },
+    },
+]
+
+
+def get_model_constructor(model_name: str):
+
+    def constructor(
+        params: dict,
+        input_shape: tuple[int, int],
+        target_type: str,
+        num_class: Optional[int] = None,
+    ):
+        """
+        Builds the recurrent model based on the initialized parameters and model name.
+        :return: A Keras Model object.
+        """
+        inputs = Input(shape=input_shape)
+
+        # Model selection logic
+        if model_name == "LSTM-1":
+            x = LSTM(**params["model_params"])(inputs)
+
+        elif model_name == "LSTM-2":
+            x = LSTM(**params["model_params"], return_sequences=True)(inputs)
+            x = LSTM(**params["model_params"])(x)
+
+        elif model_name == "LSTM-2-Deep":
+            x = LSTM(**params["model_params"], return_sequences=True)(inputs)
+            x = LSTM(**params["model_params"])(x)
+            x = Dense(50)(x)
+
+        elif model_name == "BiLSTM-1":
+            x = Bidirectional(LSTM(**params["model_params"]))(inputs)
+
+        elif model_name == "BiLSTM-2":  # TODO: create search params ?
+            x = Bidirectional(LSTM(**params["model_params"], return_sequences=True))(
+                inputs
+            )
+            x = Bidirectional(LSTM(**params["model_params"]))(x)
+
+        elif model_name == "GRU-1":
+            x = GRU(**params["model_params"])(inputs)
+
+        elif model_name == "GRU-2":  # TODO: create search params ?
+            x = GRU(**params["model_params"], return_sequences=True)(inputs)
+            x = GRU(**params["model_params"])(x)
+
+        elif model_name == "GRU-2-Deep":  # TODO: create search params ?
+            x = GRU(**params["model_params"], return_sequences=True)(inputs)
+            x = GRU(**params["model_params"])(x)
+            x = Dense(50)(x)
+
+        elif model_name == "BiGRU-1":
+            x = Bidirectional(GRU(**params["model_params"]))(inputs)
+
+        elif model_name == "BiGRU-2":  # TODO: create search params ?
+            x = Bidirectional(GRU(**params["model_params"], return_sequences=True))(
+                inputs
+            )
+            x = Bidirectional(GRU(**params["model_params"]))(x)
+
+        elif model_name == "TCN-1":
+            x = TCN(**params["model_params"])(inputs)
+
+        elif model_name == "TCN-2":  # TODO: create search params ?
+            x = TCN(**params["model_params"], return_sequences=True)(inputs)
+            x = TCN(**params["model_params"])(x)
+
+        elif model_name == "TCN-2-Deep":  # TODO: create search params ?
+            x = TCN(**params["model_params"], return_sequences=True)(inputs)
+            x = TCN(**params["model_params"])(x)
+            x = Dense(50)(x)
+
+        elif model_name == "BiTCN-1":  # TODO: create search params ?
+            x = Bidirectional(TCN(**params["model_params"], return_sequences=False))(
+                inputs
+            )
+
+        elif model_name == "BiTCN-2":  # TODO: create search params ?
+            x = Bidirectional(TCN(**params["model_params"], return_sequences=True))(
+                inputs
+            )
+            x = Bidirectional(TCN(**params["model_params"], return_sequences=False))(x)
+
+        elif model_name == "Seq2Seq":
+            # encoder
+            encoder_last_h1, encoder_last_h2, encoder_last_c = LSTM(
+                **params["model_params"], return_state=True
+            )(inputs)
+            encoder_last_h1 = LayerNormalization(epsilon=1e-6)(encoder_last_h1)
+            encoder_last_c = LayerNormalization(epsilon=1e-6)(encoder_last_c)
+
+            # decoder
+            decoder_timesteps = max(int(input_shape[0] / 5), 2)
+            decoder = RepeatVector(decoder_timesteps)(encoder_last_h1)
+            x = LSTM(**params["model_params"], return_state=False)(
+                decoder, initial_state=[encoder_last_h1, encoder_last_c]
+            )
+
+        elif model_name == "Transformer":
+
+            def transformer_encoder(
+                inputs, num_layers, head_size, num_heads, ff_dim, dropout=0
+            ):
+                for _ in range(num_layers):
+                    # Attention and Normalization
+                    x = LayerNormalization(epsilon=1e-6)(inputs)
+                    x = MultiHeadAttention(
+                        key_dim=head_size, num_heads=num_heads, dropout=dropout
+                    )(x, x)
+                    x = Add()([x, inputs])
+
+                    # Feed Forward Part
+                    y = LayerNormalization(epsilon=1e-6)(x)
+                    y = Dense(ff_dim, activation="relu")(y)
+                    y = Dropout(dropout)(y)
+                    y = Dense(inputs.shape[-1])(y)
+                    inputs = Add()([y, x])
+
+                return inputs
+
+            x = transformer_encoder(inputs, **params["model_params"])
+            x = GlobalAveragePooling1D()(x)
+            x = LayerNormalization(epsilon=1e-6)(x)
+
+        else:
+            raise ValueError(f"Invalid model name: {model_name}")
+
+        # Define output layer based on target type
+        if num_class is not None and num_class > 2:
+            outputs = Dense(
+                num_class,
+                kernel_initializer=Identity(),
+                kernel_regularizer=L2(l2=params["l2"]),
+                activation="softmax",
+            )(x)
+        else:
+            outputs = Dense(
+                1,
+                kernel_initializer=Identity(),
+                kernel_regularizer=L2(l2=params["l2"]),
+                activation=(sigmoid if target_type == "classification" else "linear"),
+            )(x)
+
+        # Build the model
+        model = Model(inputs=inputs, outputs=outputs, name=model_name)
+
+        # Set the name of the model based on its parameters
+        units = (
+            params["model_params"].get("nb_filters")
+            or params["model_params"].get("units")
+            or params["model_params"].get("head_size")
+        )
+        nb_params = model.count_params()
+        timesteps = input_shape[0]
+        model.model_name = model.name
+
+        return model
+
+    return constructor
+
+
+dl_recurrent_models = [
+    {
+        "recurrent": True,
+        "need_scaling": True,
+        "model_name": "LSTM-1",
+        "create_model": get_model_constructor("LSTM-1"),
+        "search_params": {
+            "model_params": {
+                "units": tune.choice([32, 64, 128]),
+                "activation": tune.choice(["tanh", "relu"]),
+                "recurrent_activation": tune.choice(["sigmoid", "relu"]),
+                "kernel_initializer": Identity(),
+                "recurrent_initializer": Identity(),
+                "dropout": tune.quniform(0.0, 0.5, 0.1),
+                "recurrent_dropout": tune.quniform(0.0, 0.5, 0.1),
+            },
+            "learning_rate": tune.loguniform(1e-4, 1e-2),
+            "batch_size": tune.choice([32, 64, 128]),
+            "epochs": tune.choice([50, 100, 200]),
+            "timesteps": tune.choice([5, 10, 20, 50, 120]),
+            "clipnorm": tune.quniform(0.5, 2.0, 0.5),
+            "l2": tune.loguniform(1e-6, 1e-1),
+        },
+    },
+    {
+        "recurrent": True,
+        "need_scaling": True,
+        "model_name": "LSTM-2",
+        "create_model": get_model_constructor("LSTM-2"),
+        "search_params": {
+            "model_params": {
+                "units": tune.choice([32, 64, 128]),
+                "activation": tune.choice(["tanh", "relu"]),
+                "recurrent_activation": tune.choice(["sigmoid", "relu"]),
+                "kernel_initializer": Identity(),
+                "recurrent_initializer": Identity(),
+                "dropout": tune.quniform(0.0, 0.5, 0.1),
+                "recurrent_dropout": tune.quniform(0.0, 0.5, 0.1),
+            },
+            "learning_rate": tune.loguniform(1e-4, 1e-2),
+            "batch_size": tune.choice([32, 64, 128]),
+            "epochs": tune.choice([50, 100, 200]),
+            "timesteps": tune.choice([5, 10, 20, 50, 120]),
+            "clipnorm": tune.quniform(0.5, 2.0, 0.5),
+            "l2": tune.loguniform(1e-6, 1e-1),
+        },
+    },
+    {
+        "recurrent": True,
+        "need_scaling": True,
+        "model_name": "LSTM-2-Deep",
+        "create_model": get_model_constructor("LSTM-2-Deep"),
+        "search_params": {
+            "model_params": {
+                "units": tune.choice([32, 64, 128]),
+                "activation": tune.choice(["tanh", "relu"]),
+                "recurrent_activation": tune.choice(["sigmoid", "relu"]),
+                "kernel_initializer": Identity(),
+                "recurrent_initializer": Identity(),
+                "dropout": tune.quniform(0.0, 0.5, 0.1),
+                "recurrent_dropout": tune.quniform(0.0, 0.5, 0.1),
+            },
+            "learning_rate": tune.loguniform(1e-4, 1e-2),
+            "batch_size": tune.choice([32, 64, 128]),
+            "epochs": tune.choice([50, 100, 200]),
+            "timesteps": tune.choice([5, 10, 20, 50, 120]),
+            "clipnorm": tune.quniform(0.5, 2.0, 0.5),
+            "l2": tune.loguniform(1e-6, 1e-1),
+        },
+    },
+    {
+        "recurrent": True,
+        "need_scaling": True,
+        "model_name": "BiLSTM-1",
+        "create_model": get_model_constructor("BiLSTM-1"),
+        "search_params": {
+            "model_params": {
+                "units": tune.choice([32, 64, 128]),
+                "activation": tune.choice(["tanh", "relu"]),
+                "recurrent_activation": tune.choice(["sigmoid", "relu"]),
+                "kernel_initializer": Identity(),
+                "recurrent_initializer": Identity(),
+                "dropout": tune.quniform(0.0, 0.5, 0.1),
+                "recurrent_dropout": tune.quniform(0.0, 0.5, 0.1),
+            },
+            "learning_rate": tune.loguniform(1e-4, 1e-2),
+            "batch_size": tune.choice([32, 64, 128]),
+            "epochs": tune.choice([50, 100, 200]),
+            "timesteps": tune.choice([5, 10, 20, 50, 120]),
+            "clipnorm": tune.quniform(0.5, 2.0, 0.5),
+            "l2": tune.loguniform(1e-6, 1e-1),
+        },
+    },
+    {
+        "recurrent": True,
+        "need_scaling": True,
+        "model_name": "GRU-1",
+        "create_model": get_model_constructor("GRU-1"),
+        "search_params": {
+            "model_params": {
+                "units": tune.choice([32, 64, 128]),
+                "activation": tune.choice(["tanh", "relu"]),
+                "recurrent_activation": tune.choice(["sigmoid", "relu"]),
+                "kernel_initializer": Identity(),
+                "recurrent_initializer": Identity(),
+                "dropout": tune.quniform(0.0, 0.5, 0.1),
+                "recurrent_dropout": tune.quniform(0.0, 0.5, 0.1),
+            },
+            "learning_rate": tune.loguniform(1e-4, 1e-2),
+            "batch_size": tune.choice([32, 64, 128]),
+            "epochs": tune.choice([50, 100, 200]),
+            "timesteps": tune.choice([5, 10, 20, 50, 120]),
+            "clipnorm": tune.quniform(0.5, 2.0, 0.5),
+            "l2": tune.loguniform(1e-6, 1e-1),
+        },
+    },
+    {
+        "recurrent": True,
+        "need_scaling": True,
+        "model_name": "BiGRU-1",
+        "create_model": get_model_constructor("GRU-1"),
+        "search_params": {
+            "model_params": {
+                "units": tune.choice([32, 64, 128]),
+                "activation": tune.choice(["tanh", "relu"]),
+                "recurrent_activation": tune.choice(["sigmoid", "relu"]),
+                "kernel_initializer": Identity(),
+                "recurrent_initializer": Identity(),
+                "dropout": tune.quniform(0.0, 0.5, 0.1),
+                "recurrent_dropout": tune.quniform(0.0, 0.5, 0.1),
+            },
+            "learning_rate": tune.loguniform(1e-4, 1e-2),
+            "batch_size": tune.choice([32, 64, 128]),
+            "epochs": tune.choice([50, 100, 200]),
+            "timesteps": tune.choice([5, 10, 20, 50, 120]),
+            "clipnorm": tune.quniform(0.5, 2.0, 0.5),
+            "l2": tune.loguniform(1e-6, 1e-1),
+        },
+    },
+    {
+        "recurrent": True,
+        "need_scaling": True,
+        "model_name": "TCN-1",
+        "create_model": get_model_constructor("TCN-1"),
+        "search_params": {
+            "model_params": {
+                "nb_filters": tune.choice([32, 64, 128]),
+                "kernel_size": tune.choice([2, 3, 5]),
+                "dropout_rate": tune.quniform(0.0, 0.5, 0.1),
+            },
+            "learning_rate": tune.loguniform(1e-4, 1e-2),
+            "batch_size": tune.choice([32, 64, 128]),
+            "epochs": tune.choice([50, 100, 200]),
+            "timesteps": tune.choice([5, 10, 20, 50, 120]),
+            "clipnorm": tune.quniform(0.5, 2.0, 0.5),
+            "l2": tune.loguniform(1e-6, 1e-1),
+        },
+    },
+    {
+        "recurrent": True,
+        "need_scaling": True,
+        "model_name": "Seq2Seq",
+        "create_model": get_model_constructor("Seq2Seq"),
+        "search_params": {
+            "model_params": {
+                "units": tune.choice([32, 64, 128]),
+                "kernel_initializer": Identity(),
+                "recurrent_initializer": Identity(),
+                "dropout": tune.quniform(0.0, 0.5, 0.1),
+                "recurrent_dropout": tune.quniform(0.0, 0.5, 0.1),
+            },
+            "learning_rate": tune.loguniform(1e-4, 1e-2),
+            "batch_size": tune.choice([32, 64, 128]),
+            "epochs": tune.choice([50, 100, 200]),
+            "timesteps": tune.choice([5, 10, 20, 50, 120]),
+            "clipnorm": tune.quniform(0.5, 2.0, 0.5),
+            "l2": tune.loguniform(1e-6, 1e-1),
+        },
+    },
+    {
+        "recurrent": True,
+        "need_scaling": True,
+        "model_name": "Transformer",
+        "create_model": get_model_constructor("Transformer"),
+        "search_params": {
+            "model_params": {
+                "head_size": tune.choice(
+                    [32, 64, 128, 256, 512]
+                ),  # Example of different head sizes to explore
+                "num_heads": tune.choice(
+                    [8, 16, 32]
+                ),  # Exploring different number of heads
+                "ff_dim": tune.choice(
+                    [128, 256, 512, 1024, 2048]
+                ),  # Feed-forward dimension options
+                "num_layers": tune.choice([6, 12, 24]),  # Number of transformer layers
+                "dropout": tune.quniform(
+                    0.1, 0.5, 0.1
+                ),  # Dropout rate between 0.1 and 0.5
+            },
+            "learning_rate": tune.loguniform(1e-4, 1e-2),
+            "batch_size": tune.choice([32, 64, 128]),
+            "epochs": tune.choice([50, 100, 200]),
+            "timesteps": tune.choice([5, 10, 20, 50, 120]),
+            "clipnorm": tune.quniform(0.5, 2.0, 0.5),
+            "l2": tune.loguniform(1e-6, 1e-1),
+        },
+    },
+]
+
+
+def get_models_idx(*model_names):
+    models = ml_models + dl_recurrent_models
+
+    matching_idx = [
+        i for i, model in enumerate(models) if model["model_name"] in model_names
+    ]
+    return matching_idx