birdnet-analyzer 2.0.1__py3-none-any.whl → 2.1.1__py3-none-any.whl

birdnet_analyzer/model.py

@@ -1,1212 +1,1213 @@
-# ruff: noqa: PLW0603
-"""Contains functions to use the BirdNET models."""
-
-import os
-import sys
-import warnings
-
-import numpy as np
-
-import birdnet_analyzer.config as cfg
-from birdnet_analyzer import utils
-
-SCRIPT_DIR = os.path.abspath(os.path.dirname(__file__))
-
-
-os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
-os.environ["CUDA_VISIBLE_DEVICES"] = ""
-
-warnings.filterwarnings("ignore")
-
-# Import TFLite from runtime or Tensorflow;
-# import Keras if protobuf model;
-# NOTE: we have to use TFLite if we want to use
-# the metadata model or want to extract embeddings
-try:
-    import tflite_runtime.interpreter as tflite  # type: ignore
-except ModuleNotFoundError:
-    from tensorflow import lite as tflite
-if not cfg.MODEL_PATH.endswith(".tflite"):
-    from tensorflow import keras
-
-INTERPRETER: tflite.Interpreter = None
-C_INTERPRETER: tflite.Interpreter = None
-M_INTERPRETER: tflite.Interpreter = None
-PBMODEL = None
-C_PBMODEL = None
-EMPTY_CLASS_EXCEPTION_REF = None
-
-
-def get_empty_class_exception():
-    import keras_tuner.errors
-
-    global EMPTY_CLASS_EXCEPTION_REF
-
-    if EMPTY_CLASS_EXCEPTION_REF:
-        return EMPTY_CLASS_EXCEPTION_REF
-
-    class EmptyClassException(keras_tuner.errors.FatalError):
-        """
-        Exception raised when a class is found to be empty.
-
-        Attributes:
-            index (int): The index of the empty class.
-            message (str): The error message indicating which class is empty.
-        """
-
-        def __init__(self, *args, index=None):
-            super().__init__(*args)
-            self.index = index
-            self.message = f"Class {index} is empty."
-
-    EMPTY_CLASS_EXCEPTION_REF = EmptyClassException
-    return EMPTY_CLASS_EXCEPTION_REF
-
-
-def label_smoothing(y: np.ndarray, alpha=0.1):
-    """
-    Applies label smoothing to the given labels.
-    Label smoothing is a technique used to prevent the model from becoming overconfident by adjusting the target labels.
-    It subtracts a small value (alpha) from the correct label and distributes it among the other labels.
-    Args:
-        y (numpy.ndarray): Array of labels to be smoothed. The array should be of shape (num_labels,).
-        alpha (float, optional): Smoothing parameter. Default is 0.1.
-    Returns:
-        numpy.ndarray: The smoothed labels.
-    """
-    # Subtract alpha from correct label when it is >0
-    y[y > 0] -= alpha
-
-    # Assigned alpha to all other labels
-    y[y == 0] = alpha / y.shape[0]
-
-    return y
-
-
-def mixup(x, y, augmentation_ratio=0.25, alpha=0.2):
-    """Apply mixup to the given data.
-
-    Mixup is a data augmentation technique that generates new samples by
-    mixing two samples and their labels.
-
-    Args:
-        x: Samples.
-        y: One-hot labels.
-        augmentation_ratio: The ratio of augmented samples.
-        alpha: The beta distribution parameter.
-
-    Returns:
-        Augmented data.
-    """
-    rng = np.random.default_rng(cfg.RANDOM_SEED)
-
-    # Get indices of all positive samples
-    positive_indices = np.unique(np.where(y[:, :] == 1)[0])
-
-    # Calculate the number of samples to augment based on the ratio
-    num_samples_to_augment = int(len(positive_indices) * augmentation_ratio)
-
-    # Indices of samples, that are already mixed up
-    mixed_up_indices = []
-
-    for _ in range(num_samples_to_augment):
-        # Randomly choose one instance from the positive samples
-        index = rng.choice(positive_indices)
-
-        # Choose another one, when the chosen one was already mixed up
-        while index in mixed_up_indices:
-            index = rng.choice(positive_indices)
-
-        x1, y1 = x[index], y[index]
-
-        # Randomly choose a different instance from the dataset
-        second_index = rng.choice(positive_indices)
-
-        # Choose again, when the same or an already mixed up sample was selected
-        while second_index == index or second_index in mixed_up_indices:
-            second_index = rng.choice(positive_indices)
-        x2, y2 = x[second_index], y[second_index]
-
-        # Generate a random mixing coefficient (lambda)
-        lambda_ = rng.beta(alpha, alpha)
-
-        # Mix the embeddings and labels
-        mixed_x = lambda_ * x1 + (1 - lambda_) * x2
-        mixed_y = lambda_ * y1 + (1 - lambda_) * y2
-
-        # Replace one of the original samples and labels with the augmented sample and labels
-        x[index] = mixed_x
-        y[index] = mixed_y
-
-        # Mark the sample as already mixed up
-        mixed_up_indices.append(index)
-
-    del mixed_x
-    del mixed_y
-
-    return x, y
-
-
-def random_split(x, y, val_ratio=0.2):
-    """Splits the data into training and validation data.
-
-    Makes sure that each class is represented in both sets.
-
-    Args:
-        x: Samples.
-        y: One-hot labels.
-        val_ratio: The ratio of validation data.
-
-    Returns:
-        A tuple of (x_train, y_train, x_val, y_val).
-    """
-    rng = np.random.default_rng(cfg.RANDOM_SEED)
-
-    # Get number of classes
-    num_classes = y.shape[1]
-
-    # Initialize training and validation data
-    x_train, y_train, x_val, y_val = [], [], [], []
-
-    # Split data
-    for i in range(num_classes):
-        # Get indices of positive samples of current class
-        positive_indices = np.where(y[:, i] == 1)[0]
-
-        # Get indices of negative samples of current class
-        negative_indices = np.where(y[:, i] == -1)[0]
-
-        # Get number of samples for each set
-        num_samples = len(positive_indices)
-        num_samples_train = max(1, int(num_samples * (1 - val_ratio)))
-        num_samples_val = max(0, num_samples - num_samples_train)
-
-        # Randomly choose samples for training and validation
-        rng.shuffle(positive_indices)
-        train_indices = positive_indices[:num_samples_train]
-        val_indices = positive_indices[num_samples_train : num_samples_train + num_samples_val]
-
-        # Append samples to training and validation data
-        x_train.append(x[train_indices])
-        y_train.append(y[train_indices])
-        x_val.append(x[val_indices])
-        y_val.append(y[val_indices])
-
-        # Append negative samples to training data
-        x_train.append(x[negative_indices])
-        y_train.append(y[negative_indices])
-
-    # Add samples for non-event classes to training and validation data
-    non_event_indices = np.where(np.sum(y[:, :], axis=1) == 0)[0]
-    num_samples = len(non_event_indices)
-    num_samples_train = max(1, int(num_samples * (1 - val_ratio)))
-    num_samples_val = max(0, num_samples - num_samples_train)
-    rng.shuffle(non_event_indices)
-    train_indices = non_event_indices[:num_samples_train]
-    val_indices = non_event_indices[num_samples_train : num_samples_train + num_samples_val]
-    x_train.append(x[train_indices])
-    y_train.append(y[train_indices])
-    x_val.append(x[val_indices])
-    y_val.append(y[val_indices])
-
-    # Concatenate data
-    x_train = np.concatenate(x_train)
-    y_train = np.concatenate(y_train)
-    x_val = np.concatenate(x_val)
-    y_val = np.concatenate(y_val)
-
-    # Shuffle data
-    indices = np.arange(len(x_train))
-    rng.shuffle(indices)
-    x_train = x_train[indices]
-    y_train = y_train[indices]
-
-    indices = np.arange(len(x_val))
-    rng.shuffle(indices)
-    x_val = x_val[indices]
-    y_val = y_val[indices]
-
-    return x_train, y_train, x_val, y_val
-
-
-def random_multilabel_split(x, y, val_ratio=0.2):
-    """Splits the data into training and validation data.
-
-    Makes sure that each combination of classes is represented in both sets.
-
-    Args:
-        x: Samples.
-        y: One-hot labels.
-        val_ratio: The ratio of validation data.
-
-    Returns:
-        A tuple of (x_train, y_train, x_val, y_val).
-
-    """
-    rng = np.random.default_rng(cfg.RANDOM_SEED)
-
-    # Find all combinations of labels
-    class_combinations = np.unique(y, axis=0)
-
-    # Initialize training and validation data
-    x_train, y_train, x_val, y_val = [], [], [], []
-
-    # Split the data for each combination of labels
-    for class_combination in class_combinations:
-        # find all indices
-        indices = np.where((y == class_combination).all(axis=1))[0]
-
-        # When negative sample use only for training
-        if -1 in class_combination:
-            x_train.append(x[indices])
-            y_train.append(y[indices])
-        # Otherwise split according to the validation split
-        else:
-            # Get number of samples for each set
-            num_samples = len(indices)
-            num_samples_train = max(1, int(num_samples * (1 - val_ratio)))
-            num_samples_val = max(0, num_samples - num_samples_train)
-            # Randomly choose samples for training and validation
-            rng.shuffle(indices)
-            train_indices = indices[:num_samples_train]
-            val_indices = indices[num_samples_train : num_samples_train + num_samples_val]
-            # Append samples to training and validation data
-            x_train.append(x[train_indices])
-            y_train.append(y[train_indices])
-            x_val.append(x[val_indices])
-            y_val.append(y[val_indices])
-
-    # Concatenate data
-    x_train = np.concatenate(x_train)
-    y_train = np.concatenate(y_train)
-    x_val = np.concatenate(x_val)
-    y_val = np.concatenate(y_val)
-
-    # Shuffle data
-    indices = np.arange(len(x_train))
-    rng.shuffle(indices)
-    x_train = x_train[indices]
-    y_train = y_train[indices]
-
-    indices = np.arange(len(x_val))
-    rng.shuffle(indices)
-    x_val = x_val[indices]
-    y_val = y_val[indices]
-
-    return x_train, y_train, x_val, y_val
-
-
-def upsample_core(x: np.ndarray, y: np.ndarray, min_samples: int, apply: callable, size=2):
-    """
-    Upsamples the minority class in the dataset using the specified apply function.
-    Parameters:
-        x (np.ndarray): The feature matrix.
-        y (np.ndarray): The target labels.
-        min_samples (int): The minimum number of samples required for the minority class.
-        apply (callable): A function that applies the SMOTE or any other algorithm to the data.
-        size (int, optional): The number of samples to generate in each iteration. Default is 2.
-    Returns:
-        tuple: A tuple containing the upsampled feature matrix and target labels.
-    """
-    rng = np.random.default_rng(cfg.RANDOM_SEED)
-    y_temp = []
-    x_temp = []
-
-    if cfg.BINARY_CLASSIFICATION:
-        # Determine if 1 or 0 is the minority class
-        minority_label = 1 if y.sum(axis=0) < len(y) - y.sum(axis=0) else 0
-
-        while np.where(y == minority_label)[0].shape[0] + len(y_temp) < min_samples:
-            # Randomly choose a sample from the minority class
-            random_index = rng.choice(np.where(y == minority_label)[0], size=size)
-
-            # Apply SMOTE
-            x_app, y_app = apply(x, y, random_index)
-            y_temp.append(y_app)
-            x_temp.append(x_app)
-    else:
-        for i in range(y.shape[1]):
-            while y[:, i].sum() + len(y_temp) < min_samples:
-                try:
-                    # Randomly choose a sample from the minority class
-                    random_index = rng.choice(np.where(y[:, i] == 1)[0], size=size)
-                except ValueError as e:
-                    raise get_empty_class_exception()(index=i) from e
-
-                # Apply SMOTE
-                x_app, y_app = apply(x, y, random_index)
-                y_temp.append(y_app)
-                x_temp.append(x_app)
-
-    return x_temp, y_temp
-
-
-def upsampling(x: np.ndarray, y: np.ndarray, ratio=0.5, mode="repeat"):
-    """Balance data through upsampling.
-
-    We upsample minority classes to have at least 10% (ratio=0.1) of the samples of the majority class.
-
-    Args:
-        x: Samples.
-        y: One-hot labels.
-        ratio: The minimum ratio of minority to majority samples.
-        mode: The upsampling mode. Either 'repeat', 'mean', 'linear' or 'smote'.
-
-    Returns:
-        Upsampled data.
-    """
-
-    # Set numpy random seed
-    rng = np.random.default_rng(cfg.RANDOM_SEED)
-
-    # Determine min number of samples
-    min_samples = (
-        int(max(y.sum(axis=0), len(y) - y.sum(axis=0)) * ratio)
-        if cfg.BINARY_CLASSIFICATION
-        else int(np.max(y.sum(axis=0)) * ratio)
-    )
-
-    x_temp = []
-    y_temp = []
-
-    if mode == "repeat":
-
-        def applyRepeat(x, y, random_index):
-            return x[random_index[0]], y[random_index[0]]
-
-        x_temp, y_temp = upsample_core(x, y, min_samples, applyRepeat, size=1)
-
-    elif mode == "mean":
-        # For each class with less than min_samples
-        # select two random samples and calculate the mean
-        def applyMean(x, y, random_indices):
-            # Calculate the mean of the two samples
-            mean = np.mean(x[random_indices], axis=0)
-
-            # Append the mean and label to a temp list
-            return mean, y[random_indices[0]]
-
-        x_temp, y_temp = upsample_core(x, y, min_samples, applyMean)
-
-    elif mode == "linear":
-        # For each class with less than min_samples
-        # select two random samples and calculate the linear combination
-        def applyLinearCombination(x, y, random_indices):
-            # Calculate the linear combination of the two samples
-            alpha = rng.uniform(0, 1)
-            new_sample = alpha * x[random_indices[0]] + (1 - alpha) * x[random_indices[1]]
-
-            # Append the new sample and label to a temp list
-            return new_sample, y[random_indices[0]]
-
-        x_temp, y_temp = upsample_core(x, y, min_samples, applyLinearCombination)
-
-    elif mode == "smote":
-        # For each class with less than min_samples apply SMOTE
-        def applySmote(x, y, random_index, k=5):
-            # Get the k nearest neighbors
-            distances = np.sqrt(np.sum((x - x[random_index[0]]) ** 2, axis=1))
-            indices = np.argsort(distances)[1 : k + 1]
-
-            # Randomly choose one of the neighbors
-            random_neighbor = rng.choice(indices)
-
-            # Calculate the difference vector
-            diff = x[random_neighbor] - x[random_index[0]]
-
-            # Randomly choose a weight between 0 and 1
-            weight = rng.uniform(0, 1)
-
-            # Calculate the new sample
-            new_sample = x[random_index[0]] + weight * diff
-
-            # Append the new sample and label to a temp list
-            return new_sample, y[random_index[0]]
-
-        x_temp, y_temp = upsample_core(x, y, min_samples, applySmote, size=1)
-
-    # Append the temp list to the original data
-    if len(x_temp) > 0:
-        x = np.vstack((x, np.array(x_temp)))
-        y = np.vstack((y, np.array(y_temp)))
-
-    # Shuffle data
-    indices = np.arange(len(x))
-    rng.shuffle(indices)
-    x = x[indices]
-    y = y[indices]
-
-    del x_temp
-    del y_temp
-
-    return x, y
-
-
-def save_model_params(path):
-    """Saves the model parameters to a file.
-
-    Args:
-        path: Path to the file.
-    """
-    utils.save_params(
-        path,
-        (
-            "Hidden units",
-            "Dropout",
-            "Batchsize",
-            "Learning rate",
-            "Crop mode",
-            "Crop overlap",
-            "Audio speed",
-            "Upsamling mode",
-            "Upsamling ratio",
-            "use mixup",
-            "use label smoothing",
-            "BirdNET Model version",
-        ),
-        (
-            cfg.TRAIN_HIDDEN_UNITS,
-            cfg.TRAIN_DROPOUT,
-            cfg.TRAIN_BATCH_SIZE,
-            cfg.TRAIN_LEARNING_RATE,
-            cfg.SAMPLE_CROP_MODE,
-            cfg.SIG_OVERLAP,
-            cfg.AUDIO_SPEED,
-            cfg.UPSAMPLING_MODE,
-            cfg.UPSAMPLING_RATIO,
-            cfg.TRAIN_WITH_MIXUP,
-            cfg.TRAIN_WITH_LABEL_SMOOTHING,
-            cfg.MODEL_VERSION,
-        ),
-    )
-
-
-def reset_custom_classifier():
-    """
-    Resets the custom classifier by setting the global variables C_INTERPRETER and C_PBMODEL to None.
-    This function is used to clear any existing custom classifier models and interpreters, effectively
-    resetting the state of the custom classifier.
-    """
-    global C_INTERPRETER
-    global C_PBMODEL
-
-    C_INTERPRETER = None
-    C_PBMODEL = None
-
-
-def load_model(class_output=True):
-    """
-    Loads the machine learning model based on the configuration provided.
-    This function loads either a TensorFlow Lite (TFLite) model or a protobuf model
-    depending on the file extension of the model path specified in the configuration.
-    It sets up the global variables for the model interpreter and input/output layer indices.
-
-    Args:
-        class_output (bool): If True, sets the output layer index to the classification output.
-                             If False, sets the output layer index to the feature embeddings.
-    """
-    global PBMODEL
-    global INTERPRETER
-    global INPUT_LAYER_INDEX
-    global OUTPUT_LAYER_INDEX
-
-    # Do we have to load the tflite or protobuf model?
-    if cfg.MODEL_PATH.endswith(".tflite"):
-        # Load TFLite model and allocate tensors.
-        INTERPRETER = tflite.Interpreter(
-            model_path=os.path.join(SCRIPT_DIR, cfg.MODEL_PATH), num_threads=cfg.TFLITE_THREADS
-        )
-        INTERPRETER.allocate_tensors()
-
-        # Get input and output tensors.
-        input_details = INTERPRETER.get_input_details()
-        output_details = INTERPRETER.get_output_details()
-
-        # Get input tensor index
-        INPUT_LAYER_INDEX = input_details[0]["index"]
-
-        # Get classification output or feature embeddings
-        OUTPUT_LAYER_INDEX = output_details[0]["index"] if class_output else output_details[0]["index"] - 1
-
-    else:
-        # Load protobuf model
-        # Note: This will throw a bunch of warnings about custom gradients
-        # which we will ignore until TF lets us block them
-        PBMODEL = keras.models.load_model(os.path.join(SCRIPT_DIR, cfg.MODEL_PATH), compile=False)
-
-
-def load_custom_classifier():
-    """
-    Loads a custom classifier model based on the file extension of the provided model path.
-    If the model file ends with ".tflite", it loads a TensorFlow Lite model and sets up the interpreter,
-    input layer index, output layer index, and input size.
-    If the model file does not end with ".tflite", it loads a TensorFlow SavedModel.
-    """
-    global C_INTERPRETER
-    global C_INPUT_LAYER_INDEX
-    global C_OUTPUT_LAYER_INDEX
-    global C_INPUT_SIZE
-    global C_PBMODEL
-
-    if cfg.CUSTOM_CLASSIFIER.endswith(".tflite"):
-        # Load TFLite model and allocate tensors.
-        C_INTERPRETER = tflite.Interpreter(model_path=cfg.CUSTOM_CLASSIFIER, num_threads=cfg.TFLITE_THREADS)
-        C_INTERPRETER.allocate_tensors()
-
-        # Get input and output tensors.
-        input_details = C_INTERPRETER.get_input_details()
-        output_details = C_INTERPRETER.get_output_details()
-
-        # Get input tensor index
-        C_INPUT_LAYER_INDEX = input_details[0]["index"]
-
-        C_INPUT_SIZE = input_details[0]["shape"][-1]
-
-        # Get classification output
-        C_OUTPUT_LAYER_INDEX = output_details[0]["index"]
-    else:
-        import tensorflow as tf
-
-        tf.get_logger().setLevel("ERROR")
-
-        C_PBMODEL = tf.saved_model.load(cfg.CUSTOM_CLASSIFIER)
-
-
-def load_meta_model():
-    """Loads the model for species prediction.
-
-    Initializes the model used to predict species list, based on coordinates and week of year.
-    """
-    global M_INTERPRETER
-    global M_INPUT_LAYER_INDEX
-    global M_OUTPUT_LAYER_INDEX
-
-    # Load TFLite model and allocate tensors.
-    M_INTERPRETER = tflite.Interpreter(
-        model_path=os.path.join(SCRIPT_DIR, cfg.MDATA_MODEL_PATH), num_threads=cfg.TFLITE_THREADS
-    )
-    M_INTERPRETER.allocate_tensors()
-
-    # Get input and output tensors.
-    input_details = M_INTERPRETER.get_input_details()
-    output_details = M_INTERPRETER.get_output_details()
-
-    # Get input tensor index
-    M_INPUT_LAYER_INDEX = input_details[0]["index"]
-    M_OUTPUT_LAYER_INDEX = output_details[0]["index"]
-
-
-def build_linear_classifier(num_labels, input_size, hidden_units=0, dropout=0.0):
-    """Builds a classifier.
-
-    Args:
-        num_labels: Output size.
-        input_size: Size of the input.
-        hidden_units: If > 0, creates another hidden layer with the given number of units.
-        dropout: Dropout rate.
-
-    Returns:
-        A new classifier.
-    """
-    # import keras
-    from tensorflow import keras
-
-    # Build a simple one- or two-layer linear classifier
-    model = keras.Sequential()
-
-    # Input layer
-    model.add(keras.layers.InputLayer(input_shape=(input_size,)))
-
-    # Batch normalization on input to standardize embeddings
-    model.add(keras.layers.BatchNormalization())
-
-    # Optional L2 regularization for all dense layers
-    regularizer = keras.regularizers.l2(1e-5)
-
-    # Hidden layer with improved architecture
-    if hidden_units > 0:
-        # Dropout layer before hidden layer
-        if dropout > 0:
-            model.add(keras.layers.Dropout(dropout))
-
-        # Add a hidden layer with L2 regularization
-        model.add(
-            keras.layers.Dense(
-                hidden_units, activation="relu", kernel_regularizer=regularizer, kernel_initializer="he_normal"
-            )
-        )
-
-        # Add another batch normalization after the hidden layer
-        model.add(keras.layers.BatchNormalization())
-
-    # Dropout layer before output
-    if dropout > 0:
-        model.add(keras.layers.Dropout(dropout))
-
-    # Classification layer with L2 regularization
-    model.add(keras.layers.Dense(num_labels, kernel_regularizer=regularizer, kernel_initializer="glorot_uniform"))
-
-    # Activation layer
-    model.add(keras.layers.Activation("sigmoid"))
-
-    return model
-
-
-def train_linear_classifier(
-    classifier,
-    x_train,
-    y_train,
-    x_test,
-    y_test,
-    epochs,
-    batch_size,
-    learning_rate,
-    val_split,
-    upsampling_ratio,
-    upsampling_mode,
-    train_with_mixup,
-    train_with_label_smoothing,
-    train_with_focal_loss=False,
-    focal_loss_gamma=2.0,
-    focal_loss_alpha=0.25,
-    on_epoch_end=None,
-):
-    """Trains a custom classifier.
-
-    Trains a new classifier for BirdNET based on the given data.
-
-    Args:
-        classifier: The classifier to be trained.
-        x_train: Samples.
-        y_train: Labels.
-        x_test: Validation samples.
-        y_test: Validation labels.
-        epochs: Number of epochs to train.
-        batch_size: Batch size.
-        learning_rate: The learning rate during training.
-        val_split: Validation split ratio (is 0 when using test data).
-        upsampling_ratio: Upsampling ratio.
-        upsampling_mode: Upsampling mode.
-        train_with_mixup: If True, applies mixup to the training data.
-        train_with_label_smoothing: If True, applies label smoothing to the training data.
-        train_with_focal_loss: If True, uses focal loss instead of binary cross-entropy loss.
-        focal_loss_gamma: Focal loss gamma parameter.
-        focal_loss_alpha: Focal loss alpha parameter.
-        on_epoch_end: Optional callback `function(epoch, logs)`.
-
-    Returns:
-        (classifier, history)
-    """
-    # import keras
-    from tensorflow import keras
-
-    class FunctionCallback(keras.callbacks.Callback):
-        def __init__(self, on_epoch_end=None) -> None:
-            super().__init__()
-            self.on_epoch_end_fn = on_epoch_end
-
-        def on_epoch_end(self, epoch, logs=None):
-            if self.on_epoch_end_fn:
-                self.on_epoch_end_fn(epoch, logs)
-
-    # Set random seed
-    rng = np.random.default_rng(cfg.RANDOM_SEED)
-
-    # Shuffle data
-    idx = np.arange(x_train.shape[0])
-    rng.shuffle(idx)
-    x_train = x_train[idx]
-    y_train = y_train[idx]
-
-    # Random val split
-    if val_split > 0:
-        if not cfg.MULTI_LABEL:
-            x_train, y_train, x_val, y_val = random_split(x_train, y_train, val_split)
-        else:
-            x_train, y_train, x_val, y_val = random_multilabel_split(x_train, y_train, val_split)
-    else:
-        x_val = x_test
-        y_val = y_test
-
-    print(
-        f"Training on {x_train.shape[0]} samples, validating on {x_val.shape[0]} samples.",
-        flush=True,
-    )
-
-    # Upsample training data
-    if upsampling_ratio > 0:
-        x_train, y_train = upsampling(x_train, y_train, upsampling_ratio, upsampling_mode)
-        print(f"Upsampled training data to {x_train.shape[0]} samples.", flush=True)
-
-    # Apply mixup to training data
-    if train_with_mixup and not cfg.BINARY_CLASSIFICATION:
-        x_train, y_train = mixup(x_train, y_train)
-
-    # Apply label smoothing
-    if train_with_label_smoothing and not cfg.BINARY_CLASSIFICATION:
-        y_train = label_smoothing(y_train)
-
-    # Early stopping with patience depending on dataset size
-    patience = min(10, max(5, int(epochs / 10)))
-    min_delta = 0.001
-
-    callbacks = [
-        # EarlyStopping with restore_best_weights
-        keras.callbacks.EarlyStopping(
-            monitor="val_AUPRC",
-            mode="max",
-            patience=patience,
-            verbose=1,
-            min_delta=min_delta,
-            restore_best_weights=True,
-        ),
-        # Function callback for progress tracking
-        FunctionCallback(on_epoch_end=on_epoch_end),
-    ]
-
-    # Learning rate schedule - use cosine decay with warmup
-    warmup_epochs = min(5, int(epochs * 0.1))
-
-    def lr_schedule(epoch, lr):
-        if epoch < warmup_epochs:
-            # Linear warmup
-            return learning_rate * (epoch + 1) / warmup_epochs
-
-        # Cosine decay
-        progress = (epoch - warmup_epochs) / (epochs - warmup_epochs)
-        return learning_rate * (0.1 + 0.9 * (1 + np.cos(np.pi * progress)) / 2)
-
-    # Add LR scheduler callback
-    callbacks.append(keras.callbacks.LearningRateScheduler(lr_schedule))
-
-    optimizer_cls = keras.optimizers.legacy.Adam if sys.platform == "darwin" else keras.optimizers.Adam
-
-    def _focal_loss(y_true, y_pred):
-        return focal_loss(y_true, y_pred, gamma=cfg.FOCAL_LOSS_GAMMA, alpha=cfg.FOCAL_LOSS_ALPHA)
-
-    # Choose the loss function based on config
-    loss_function = _focal_loss if train_with_focal_loss else custom_loss
-
-    # Compile model with appropriate metrics for classification task
-    classifier.compile(
-        optimizer=optimizer_cls(learning_rate=learning_rate),
-        loss=loss_function,
-        metrics=[
-            keras.metrics.AUC(
-                curve="PR",
-                multi_label=cfg.MULTI_LABEL,
-                name="AUPRC",
-                num_labels=y_train.shape[1] if cfg.MULTI_LABEL else None,
-                from_logits=True,
-            ),
-            keras.metrics.AUC(
-                curve="ROC",
-                multi_label=cfg.MULTI_LABEL,
-                name="AUROC",
-                num_labels=y_train.shape[1] if cfg.MULTI_LABEL else None,
-                from_logits=True,
-            ),
-        ],
-    )
-
-    # Train model
-    history = classifier.fit(
-        x_train, y_train, epochs=epochs, batch_size=batch_size, validation_data=(x_val, y_val), callbacks=callbacks
-    )
-
-    return classifier, history
-
-
-def save_linear_classifier(classifier, model_path: str, labels: list[str], mode="replace"):
-    """Saves the classifier as a tflite model, as well as the used labels in a .txt.
-
-    Args:
-        classifier: The custom classifier.
-        model_path: Path the model will be saved at.
-        labels: List of labels used for the classifier.
-    """
-    import tensorflow as tf
-
-    global PBMODEL
-
-    tf.get_logger().setLevel("ERROR")
-
-    if PBMODEL is None:
-        PBMODEL = tf.keras.models.load_model(os.path.join(SCRIPT_DIR, cfg.PB_MODEL), compile=False)
-
-    saved_model = PBMODEL
-
-    # Remove activation layer
-    classifier.pop()
-
-    if mode == "replace":
-        combined_model = tf.keras.Sequential([saved_model.embeddings_model, classifier], "basic")
-    elif mode == "append":
-        intermediate = classifier(saved_model.model.get_layer("GLOBAL_AVG_POOL").output)
-
-        output = tf.keras.layers.concatenate([saved_model.model.output, intermediate], name="combined_output")
-
-        combined_model = tf.keras.Model(inputs=saved_model.model.input, outputs=output)
-    else:
-        raise ValueError("Model save mode must be either 'replace' or 'append'")
-
-    # Append .tflite if necessary
-    if not model_path.endswith(".tflite"):
-        model_path += ".tflite"
-
-    # Make folders
-    os.makedirs(os.path.dirname(model_path), exist_ok=True)
-
-    # Save model as tflite
-    converter = tf.lite.TFLiteConverter.from_keras_model(combined_model)
-    tflite_model = converter.convert()
-
-    with open(model_path, "wb") as f:
-        f.write(tflite_model)
-
-    if mode == "append":
-        labels = [*utils.read_lines(os.path.join(SCRIPT_DIR, cfg.LABELS_FILE)), *labels]
-
-    # Save labels
-    with open(model_path.replace(".tflite", "_Labels.txt"), "w", encoding="utf-8") as f:
-        for label in labels:
-            f.write(label + "\n")
-
-    save_model_params(model_path.replace(".tflite", "_Params.csv"))
-
-
-def save_raven_model(classifier, model_path: str, labels: list[str], mode="replace"):
-    """
-    Save a TensorFlow model with a custom classifier and associated metadata for use with BirdNET.
-
-    Args:
-        classifier (tf.keras.Model): The custom classifier model to be saved.
-        model_path (str): The path where the model will be saved.
-        labels (list[str]): A list of labels associated with the classifier.
-        mode (str, optional): The mode for saving the model. Can be either "replace" or "append".
-                              Defaults to "replace".
-
-    Raises:
-        ValueError: If the mode is not "replace" or "append".
-
-    Returns:
-        None
-    """
-    import csv
-    import json
-
-    import tensorflow as tf
-
-    global PBMODEL
-
-    tf.get_logger().setLevel("ERROR")
-
-    if PBMODEL is None:
-        PBMODEL = tf.keras.models.load_model(os.path.join(SCRIPT_DIR, cfg.PB_MODEL), compile=False)
-
-    saved_model = PBMODEL
-
-    if mode == "replace":
-        combined_model = tf.keras.Sequential([saved_model.embeddings_model, classifier], "basic")
-    elif mode == "append":
-        # Remove activation layer
-        classifier.pop()
-        intermediate = classifier(saved_model.model.get_layer("GLOBAL_AVG_POOL").output)
-
-        output = tf.keras.layers.concatenate([saved_model.model.output, intermediate], name="combined_output")
-
-        combined_model = tf.keras.Model(inputs=saved_model.model.input, outputs=output)
-    else:
-        raise ValueError("Model save mode must be either 'replace' or 'append'")
-
-    # Make signatures
-    class SignatureModule(tf.Module):
-        def __init__(self, keras_model):
-            super().__init__()
-            self.model = keras_model
-
-        @tf.function(input_signature=[tf.TensorSpec(shape=[None, 144000], dtype=tf.float32)])
-        def basic(self, inputs):
-            return {"scores": self.model(inputs)}
-
-    smodel = SignatureModule(combined_model)
-    signatures = {
-        "basic": smodel.basic,
-    }
-
-    # Save signature model
-    os.makedirs(os.path.dirname(model_path), exist_ok=True)
-    model_path = model_path.removesuffix(".tflite")
-    tf.saved_model.save(smodel, model_path, signatures=signatures)
-
-    if mode == "append":
-        labels = [*utils.read_lines(os.path.join(SCRIPT_DIR, cfg.LABELS_FILE)), *labels]
-
-    # Save label file
-    labelIds = [label[:4].replace(" ", "") + str(i) for i, label in enumerate(labels, 1)]
-    labels_dir = os.path.join(model_path, "labels")
-
-    os.makedirs(labels_dir, exist_ok=True)
-
-    with open(os.path.join(labels_dir, "label_names.csv"), "w", newline="") as labelsfile:
-        labelwriter = csv.writer(labelsfile)
-        labelwriter.writerows(zip(labelIds, labels, strict=True))
-
-    # Save class names file
-    classes_dir = os.path.join(model_path, "classes")
-
-    os.makedirs(classes_dir, exist_ok=True)
-
-    with open(os.path.join(classes_dir, "classes.csv"), "w", newline="") as classesfile:
-        classeswriter = csv.writer(classesfile)
-        for labelId in labelIds:
-            classeswriter.writerow((labelId, 0.25, cfg.SIG_FMIN, cfg.SIG_FMAX, False))
-
-    # Save model config
-    model_config = os.path.join(model_path, "model_config.json")
-
-    with open(model_config, "w") as modelconfigfile:
-        modelconfig = {
-            "specVersion": 1,
-            "modelDescription": "Custom classifier trained with BirdNET "
-            + cfg.MODEL_VERSION
-            + " embeddings.\n"
-            + "BirdNET was developed by the K. Lisa Yang Center for Conservation Bioacoustics"
-            + "at the Cornell Lab of Ornithology in collaboration with Chemnitz University of Technology.\n\n"
-            + "https://birdnet.cornell.edu",
-            "modelTypeConfig": {"modelType": "RECOGNITION"},
-            "signatures": [
-                {
-                    "signatureName": "basic",
-                    "modelInputs": [
-                        {
-                            "inputName": "inputs",
-                            "sampleRate": 48000.0,
-                            "inputConfig": ["batch", "samples"],
-                        }
-                    ],
-                    "modelOutputs": [{"outputName": "scores", "outputType": "SCORES"}],
-                }
-            ],
-            "globalSemanticKeys": labelIds,
-        }
-        json.dump(modelconfig, modelconfigfile, indent=2)
-
-        model_params = os.path.join(model_path, "model_params.csv")
-
-        save_model_params(model_params)
-
-
-def predict_filter(lat, lon, week):
-    """Predicts the probability for each species.
-
-    Args:
-        lat: The latitude.
-        lon: The longitude.
-        week: The week of the year [1-48]. Use -1 for yearlong.
-
-    Returns:
-        A list of probabilities for all species.
-    """
-    # Does interpreter exist?
-    if M_INTERPRETER is None:
-        load_meta_model()
-
-    # Prepare mdata as sample
-    sample = np.expand_dims(np.array([lat, lon, week], dtype="float32"), 0)
-
-    # Run inference
-    M_INTERPRETER.set_tensor(M_INPUT_LAYER_INDEX, sample)
-    M_INTERPRETER.invoke()
-
-    return M_INTERPRETER.get_tensor(M_OUTPUT_LAYER_INDEX)[0]
-
-
-def explore(lat: float, lon: float, week: int):
-    """Predicts the species list.
-
-    Predicts the species list based on the coordinates and week of year.
-
-    Args:
-        lat: The latitude.
-        lon: The longitude.
-        week: The week of the year [1-48]. Use -1 for yearlong.
-
-    Returns:
-        A sorted list of tuples with the score and the species.
-    """
-    # Make filter prediction
-    l_filter = predict_filter(lat, lon, week)
-
-    # Apply threshold
-    l_filter = np.where(l_filter >= cfg.LOCATION_FILTER_THRESHOLD, l_filter, 0)
-
-    # Zip with labels
-    l_filter = list(zip(l_filter, cfg.LABELS, strict=True))
-
-    # Sort by filter value
-    return sorted(l_filter, key=lambda x: x[0], reverse=True)
-
-
-def focal_loss(y_true, y_pred, gamma=2.0, alpha=0.25, epsilon=1e-7):
-    """
-    Focal loss for better handling of class imbalance.
-
-    This loss function gives more weight to hard examples and down-weights easy examples.
-    Particularly helpful for imbalanced datasets where some classes have few samples.
-
-    Args:
-        y_true: Ground truth labels.
-        y_pred: Predicted probabilities.
-        gamma: Focusing parameter. Higher values mean more focus on hard examples.
-        alpha: Balance parameter. Controls weight of positive vs negative examples.
-        epsilon: Small constant to prevent log(0).
-
-    Returns:
-        Focal loss value.
-    """
-    import tensorflow.keras.backend as K
-
-    # Apply sigmoid if not already applied
-    y_pred = K.clip(y_pred, epsilon, 1.0 - epsilon)
-
-    # Calculate cross entropy
-    cross_entropy = -y_true * K.log(y_pred) - (1 - y_true) * K.log(1 - y_pred)
-
-    # Calculate focal weight
-    p_t = y_true * y_pred + (1 - y_true) * (1 - y_pred)
-    focal_weight = K.pow(1 - p_t, gamma)
-
-    # Apply alpha balancing
-    alpha_factor = y_true * alpha + (1 - y_true) * (1 - alpha)
-
-    # Calculate focal loss
-    focal_loss = alpha_factor * focal_weight * cross_entropy
-
-    # Sum over all classes
-    return K.sum(focal_loss, axis=-1)
-
-
-def custom_loss(y_true, y_pred, epsilon=1e-7):
-    import tensorflow.keras.backend as K
-
-    # Calculate loss for positive labels with epsilon
-    positive_loss = -K.sum(y_true * K.log(K.clip(y_pred, epsilon, 1.0 - epsilon)), axis=-1)
-
-    # Calculate loss for negative labels with epsilon
-    negative_loss = -K.sum((1 - y_true) * K.log(K.clip(1 - y_pred, epsilon, 1.0 - epsilon)), axis=-1)
-
-    # Combine both loss terms
-    return positive_loss + negative_loss
-
-
-def flat_sigmoid(x, sensitivity=-1, bias=1.0):
-    """
-    Applies a flat sigmoid function to the input array with a bias shift.
-
-    The flat sigmoid function is defined as:
-        f(x) = 1 / (1 + exp(sensitivity * clip(x + bias, -20, 20)))
-
-    We transform the bias parameter to a range of [-100, 100] with the formula:
-        transformed_bias = (bias - 1.0) * 10.0
-
-    Thus, higher bias values will shift the sigmoid function to the right on the x-axis, making it more "sensitive".
-
-    Note: Not sure why we are clipping, must be for numerical stability somewhere else in the code.
-
-    Args:
-        x (array-like): Input data.
-        sensitivity (float, optional): Sensitivity parameter for the sigmoid function. Default is -1.
-        bias (float, optional): Bias parameter to shift the sigmoid function on the x-axis. Must be in the range [0.01, 1.99]. Default is 1.0.
-
-    Returns:
-        numpy.ndarray: Transformed data after applying the flat sigmoid function.
-    """
-
-    transformed_bias = (bias - 1.0) * 10.0
-
-    return 1 / (1.0 + np.exp(sensitivity * np.clip(x + transformed_bias, -20, 20)))
-
-
-def predict(sample):
-    """Uses the main net to predict a sample.
-
-    Args:
-        sample: Audio sample.
-
-    Returns:
-        The prediction scores for the sample.
-    """
-    # Has custom classifier?
-    if cfg.CUSTOM_CLASSIFIER is not None:
-        return predict_with_custom_classifier(sample)
-
-    # Does interpreter or keras model exist?
-    if INTERPRETER is None and PBMODEL is None:
-        load_model()
-
-    if PBMODEL is None:
-        # Reshape input tensor
-        INTERPRETER.resize_tensor_input(INPUT_LAYER_INDEX, [len(sample), *sample[0].shape])
-        INTERPRETER.allocate_tensors()
-
-        # Make a prediction (Audio only for now)
-        INTERPRETER.set_tensor(INPUT_LAYER_INDEX, np.array(sample, dtype="float32"))
-        INTERPRETER.invoke()
-        return INTERPRETER.get_tensor(OUTPUT_LAYER_INDEX)
-
-    # Make a prediction (Audio only for now)
-    return PBMODEL.basic(sample)["scores"]
-
-
-def predict_with_custom_classifier(sample):
-    """Uses the custom classifier to make a prediction.
-
-    Args:
-        sample: Audio sample.
-
-    Returns:
-        The prediction scores for the sample.
-    """
-    # Does interpreter exist?
-    if C_INTERPRETER is None and C_PBMODEL is None:
-        load_custom_classifier()
-
-    if C_PBMODEL is None:
-        vector = embeddings(sample) if C_INPUT_SIZE != 144000 else sample
-
-        # Reshape input tensor
-        C_INTERPRETER.resize_tensor_input(C_INPUT_LAYER_INDEX, [len(vector), *vector[0].shape])
-        C_INTERPRETER.allocate_tensors()
-
-        # Make a prediction
-        C_INTERPRETER.set_tensor(C_INPUT_LAYER_INDEX, np.array(vector, dtype="float32"))
-        C_INTERPRETER.invoke()
-
-        return C_INTERPRETER.get_tensor(C_OUTPUT_LAYER_INDEX)
-
-    return C_PBMODEL.basic(sample)["scores"]
-
-
-def embeddings(sample):
-    """Extracts the embeddings for a sample.
-
-    Args:
-        sample: Audio samples.
-
-    Returns:
-        The embeddings.
-    """
-    # Does interpreter exist?
-    if INTERPRETER is None:
-        load_model(False)
-
-    # Reshape input tensor
-    INTERPRETER.resize_tensor_input(INPUT_LAYER_INDEX, [len(sample), *sample[0].shape])
-    INTERPRETER.allocate_tensors()
-
-    # Extract feature embeddings
-    INTERPRETER.set_tensor(INPUT_LAYER_INDEX, np.array(sample, dtype="float32"))
-    INTERPRETER.invoke()
-
-    return INTERPRETER.get_tensor(OUTPUT_LAYER_INDEX)
+# ruff: noqa: PLW0603
+"""Contains functions to use the BirdNET models."""
+
+import os
+import sys
+import warnings
+
+import numpy as np
+
+import birdnet_analyzer.config as cfg
+from birdnet_analyzer import utils
+
+SCRIPT_DIR = os.path.abspath(os.path.dirname(__file__))
+
+
+os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
+os.environ["CUDA_VISIBLE_DEVICES"] = ""
+
+warnings.filterwarnings("ignore")
+
+# Import TFLite from runtime or Tensorflow;
+# import Keras if protobuf model;
+# NOTE: we have to use TFLite if we want to use
+# the metadata model or want to extract embeddings
+try:
+    import tflite_runtime.interpreter as tflite  # type: ignore
+except ModuleNotFoundError:
+    from tensorflow import lite as tflite
+if not cfg.MODEL_PATH.endswith(".tflite"):
+    from tensorflow import keras
+
+INTERPRETER: tflite.Interpreter = None
+C_INTERPRETER: tflite.Interpreter = None
+M_INTERPRETER: tflite.Interpreter = None
+OUTPUT_DETAILS = None
+PBMODEL = None
+C_PBMODEL = None
+EMPTY_CLASS_EXCEPTION_REF = None
+
+
+def get_empty_class_exception():
+    import keras_tuner.errors
+
+    global EMPTY_CLASS_EXCEPTION_REF
+
+    if EMPTY_CLASS_EXCEPTION_REF:
+        return EMPTY_CLASS_EXCEPTION_REF
+
+    class EmptyClassException(keras_tuner.errors.FatalError):
+        """
+        Exception raised when a class is found to be empty.
+
+        Attributes:
+            index (int): The index of the empty class.
+            message (str): The error message indicating which class is empty.
+        """
+
+        def __init__(self, *args, index=None):
+            super().__init__(*args)
+            self.index = index
+            self.message = f"Class {index} is empty."
+
+    EMPTY_CLASS_EXCEPTION_REF = EmptyClassException
+    return EMPTY_CLASS_EXCEPTION_REF
+
+
+def label_smoothing(y: np.ndarray, alpha=0.1):
+    """
+    Applies label smoothing to the given labels.
+    Label smoothing is a technique used to prevent the model from becoming overconfident by adjusting the target labels.
+    It subtracts a small value (alpha) from the correct label and distributes it among the other labels.
+    Args:
+        y (numpy.ndarray): Array of labels to be smoothed. The array should be of shape (num_labels,).
+        alpha (float, optional): Smoothing parameter. Default is 0.1.
+    Returns:
+        numpy.ndarray: The smoothed labels.
+    """
+    # Subtract alpha from correct label when it is >0
+    y[y > 0] -= alpha
+
+    # Assigned alpha to all other labels
+    y[y == 0] = alpha / y.shape[0]
+
+    return y
+
+
+def mixup(x, y, augmentation_ratio=0.25, alpha=0.2):
+    """Apply mixup to the given data.
+
+    Mixup is a data augmentation technique that generates new samples by
+    mixing two samples and their labels.
+
+    Args:
+        x: Samples.
+        y: One-hot labels.
+        augmentation_ratio: The ratio of augmented samples.
+        alpha: The beta distribution parameter.
+
+    Returns:
+        Augmented data.
+    """
+    rng = np.random.default_rng(cfg.RANDOM_SEED)
+
+    # Get indices of all positive samples
+    positive_indices = np.unique(np.where(y[:, :] == 1)[0])
+
+    # Calculate the number of samples to augment based on the ratio
+    num_samples_to_augment = int(len(positive_indices) * augmentation_ratio)
+
+    # Indices of samples, that are already mixed up
+    mixed_up_indices = []
+
+    for _ in range(num_samples_to_augment):
+        # Randomly choose one instance from the positive samples
+        index = rng.choice(positive_indices)
+
+        # Choose another one, when the chosen one was already mixed up
+        while index in mixed_up_indices:
+            index = rng.choice(positive_indices)
+
+        x1, y1 = x[index], y[index]
+
+        # Randomly choose a different instance from the dataset
+        second_index = rng.choice(positive_indices)
+
+        # Choose again, when the same or an already mixed up sample was selected
+        while second_index == index or second_index in mixed_up_indices:
+            second_index = rng.choice(positive_indices)
+        x2, y2 = x[second_index], y[second_index]
+
+        # Generate a random mixing coefficient (lambda)
+        lambda_ = rng.beta(alpha, alpha)
+
+        # Mix the embeddings and labels
+        mixed_x = lambda_ * x1 + (1 - lambda_) * x2
+        mixed_y = lambda_ * y1 + (1 - lambda_) * y2
+
+        # Replace one of the original samples and labels with the augmented sample and labels
+        x[index] = mixed_x
+        y[index] = mixed_y
+
+        # Mark the sample as already mixed up
+        mixed_up_indices.append(index)
+
+        del mixed_x
+        del mixed_y
+
+    return x, y
+
+
+def random_split(x, y, val_ratio=0.2):
+    """Splits the data into training and validation data.
+
+    Makes sure that each class is represented in both sets.
+
+    Args:
+        x: Samples.
+        y: One-hot labels.
+        val_ratio: The ratio of validation data.
+
+    Returns:
+        A tuple of (x_train, y_train, x_val, y_val).
+    """
+    rng = np.random.default_rng(cfg.RANDOM_SEED)
+
+    # Get number of classes
+    num_classes = y.shape[1]
+
+    # Initialize training and validation data
+    x_train, y_train, x_val, y_val = [], [], [], []
+
+    # Split data
+    for i in range(num_classes):
+        # Get indices of positive samples of current class
+        positive_indices = np.where(y[:, i] == 1)[0]
+
+        # Get indices of negative samples of current class
+        negative_indices = np.where(y[:, i] == -1)[0]
+
+        # Get number of samples for each set
+        num_samples = len(positive_indices)
+        num_samples_train = max(1, int(num_samples * (1 - val_ratio)))
+        num_samples_val = max(0, num_samples - num_samples_train)
+
+        # Randomly choose samples for training and validation
+        rng.shuffle(positive_indices)
+        train_indices = positive_indices[:num_samples_train]
+        val_indices = positive_indices[num_samples_train : num_samples_train + num_samples_val]
+
+        # Append samples to training and validation data
+        x_train.append(x[train_indices])
+        y_train.append(y[train_indices])
+        x_val.append(x[val_indices])
+        y_val.append(y[val_indices])
+
+        # Append negative samples to training data
+        x_train.append(x[negative_indices])
+        y_train.append(y[negative_indices])
+
+    # Add samples for non-event classes to training and validation data
+    non_event_indices = np.where(np.sum(y[:, :], axis=1) == 0)[0]
+    num_samples = len(non_event_indices)
+    num_samples_train = max(1, int(num_samples * (1 - val_ratio)))
+    num_samples_val = max(0, num_samples - num_samples_train)
+    rng.shuffle(non_event_indices)
+    train_indices = non_event_indices[:num_samples_train]
+    val_indices = non_event_indices[num_samples_train : num_samples_train + num_samples_val]
+    x_train.append(x[train_indices])
+    y_train.append(y[train_indices])
+    x_val.append(x[val_indices])
+    y_val.append(y[val_indices])
+
+    # Concatenate data
+    x_train = np.concatenate(x_train)
+    y_train = np.concatenate(y_train)
+    x_val = np.concatenate(x_val)
+    y_val = np.concatenate(y_val)
+
+    # Shuffle data
+    indices = np.arange(len(x_train))
+    rng.shuffle(indices)
+    x_train = x_train[indices]
+    y_train = y_train[indices]
+
+    indices = np.arange(len(x_val))
+    rng.shuffle(indices)
+    x_val = x_val[indices]
+    y_val = y_val[indices]
+
+    return x_train, y_train, x_val, y_val
+
+
+def random_multilabel_split(x, y, val_ratio=0.2):
+    """Splits the data into training and validation data.
+
+    Makes sure that each combination of classes is represented in both sets.
+
+    Args:
+        x: Samples.
+        y: One-hot labels.
+        val_ratio: The ratio of validation data.
+
+    Returns:
+        A tuple of (x_train, y_train, x_val, y_val).
+
+    """
+    rng = np.random.default_rng(cfg.RANDOM_SEED)
+
+    # Find all combinations of labels
+    class_combinations = np.unique(y, axis=0)
+
+    # Initialize training and validation data
+    x_train, y_train, x_val, y_val = [], [], [], []
+
+    # Split the data for each combination of labels
+    for class_combination in class_combinations:
+        # find all indices
+        indices = np.where((y == class_combination).all(axis=1))[0]
+
+        # When negative sample use only for training
+        if -1 in class_combination:
+            x_train.append(x[indices])
+            y_train.append(y[indices])
+        # Otherwise split according to the validation split
+        else:
+            # Get number of samples for each set
+            num_samples = len(indices)
+            num_samples_train = max(1, int(num_samples * (1 - val_ratio)))
+            num_samples_val = max(0, num_samples - num_samples_train)
+            # Randomly choose samples for training and validation
+            rng.shuffle(indices)
+            train_indices = indices[:num_samples_train]
+            val_indices = indices[num_samples_train : num_samples_train + num_samples_val]
+            # Append samples to training and validation data
+            x_train.append(x[train_indices])
+            y_train.append(y[train_indices])
+            x_val.append(x[val_indices])
+            y_val.append(y[val_indices])
+
+    # Concatenate data
+    x_train = np.concatenate(x_train)
+    y_train = np.concatenate(y_train)
+    x_val = np.concatenate(x_val)
+    y_val = np.concatenate(y_val)
+
+    # Shuffle data
+    indices = np.arange(len(x_train))
+    rng.shuffle(indices)
+    x_train = x_train[indices]
+    y_train = y_train[indices]
+
+    indices = np.arange(len(x_val))
+    rng.shuffle(indices)
+    x_val = x_val[indices]
+    y_val = y_val[indices]
+
+    return x_train, y_train, x_val, y_val
+
+
+def upsample_core(x: np.ndarray, y: np.ndarray, min_samples: int, apply, size=2):
+    """
+    Upsamples the minority class in the dataset using the specified apply function.
+    Parameters:
+        x (np.ndarray): The feature matrix.
+        y (np.ndarray): The target labels.
+        min_samples (int): The minimum number of samples required for the minority class.
+        apply (callable): A function that applies the SMOTE or any other algorithm to the data.
+        size (int, optional): The number of samples to generate in each iteration. Default is 2.
+    Returns:
+        tuple: A tuple containing the upsampled feature matrix and target labels.
+    """
+    rng = np.random.default_rng(cfg.RANDOM_SEED)
+    y_temp = []
+    x_temp = []
+
+    if cfg.BINARY_CLASSIFICATION:
+        # Determine if 1 or 0 is the minority class
+        minority_label = 1 if y.sum(axis=0) < len(y) - y.sum(axis=0) else 0
+
+        while np.where(y == minority_label)[0].shape[0] + len(y_temp) < min_samples:
+            # Randomly choose a sample from the minority class
+            random_index = rng.choice(np.where(y == minority_label)[0], size=size)
+
+            # Apply SMOTE
+            x_app, y_app = apply(x, y, random_index)
+            y_temp.append(y_app)
+            x_temp.append(x_app)
+    else:
+        for i in range(y.shape[1]):
+            while y[:, i].sum() + len(y_temp) < min_samples:
+                try:
+                    # Randomly choose a sample from the minority class
+                    random_index = rng.choice(np.where(y[:, i] == 1)[0], size=size)
+                except ValueError as e:
+                    raise get_empty_class_exception()(index=i) from e
+
+                # Apply SMOTE
+                x_app, y_app = apply(x, y, random_index)
+                y_temp.append(y_app)
+                x_temp.append(x_app)
+
+    return x_temp, y_temp
+
+
+def upsampling(x: np.ndarray, y: np.ndarray, ratio=0.5, mode="repeat"):
+    """Balance data through upsampling.
+
+    We upsample minority classes to have at least 10% (ratio=0.1) of the samples of the majority class.
+
+    Args:
+        x: Samples.
+        y: One-hot labels.
+        ratio: The minimum ratio of minority to majority samples.
+        mode: The upsampling mode. Either 'repeat', 'mean', 'linear' or 'smote'.
+
+    Returns:
+        Upsampled data.
+    """
+
+    # Set numpy random seed
+    rng = np.random.default_rng(cfg.RANDOM_SEED)
+
+    # Determine min number of samples
+    min_samples = (
+        int(max(y.sum(axis=0), len(y) - y.sum(axis=0)) * ratio)
+        if cfg.BINARY_CLASSIFICATION
+        else int(np.max(y.sum(axis=0)) * ratio)
+    )
+
+    x_temp = []
+    y_temp = []
+
+    if mode == "repeat":
+
+        def applyRepeat(x, y, random_index):
+            return x[random_index[0]], y[random_index[0]]
+
+        x_temp, y_temp = upsample_core(x, y, min_samples, applyRepeat, size=1)
+
+    elif mode == "mean":
+        # For each class with less than min_samples
+        # select two random samples and calculate the mean
+        def applyMean(x, y, random_indices):
+            # Calculate the mean of the two samples
+            mean = np.mean(x[random_indices], axis=0)
+
+            # Append the mean and label to a temp list
+            return mean, y[random_indices[0]]
+
+        x_temp, y_temp = upsample_core(x, y, min_samples, applyMean)
+
+    elif mode == "linear":
+        # For each class with less than min_samples
+        # select two random samples and calculate the linear combination
+        def applyLinearCombination(x, y, random_indices):
+            # Calculate the linear combination of the two samples
+            alpha = rng.uniform(0, 1)
+            new_sample = alpha * x[random_indices[0]] + (1 - alpha) * x[random_indices[1]]
+
+            # Append the new sample and label to a temp list
+            return new_sample, y[random_indices[0]]
+
+        x_temp, y_temp = upsample_core(x, y, min_samples, applyLinearCombination)
+
+    elif mode == "smote":
+        # For each class with less than min_samples apply SMOTE
+        def applySmote(x, y, random_index, k=5):
+            # Get the k nearest neighbors
+            distances = np.sqrt(np.sum((x - x[random_index[0]]) ** 2, axis=1))
+            indices = np.argsort(distances)[1 : k + 1]
+
+            # Randomly choose one of the neighbors
+            random_neighbor = rng.choice(indices)
+
+            # Calculate the difference vector
+            diff = x[random_neighbor] - x[random_index[0]]
+
+            # Randomly choose a weight between 0 and 1
+            weight = rng.uniform(0, 1)
+
+            # Calculate the new sample
+            new_sample = x[random_index[0]] + weight * diff
+
+            # Append the new sample and label to a temp list
+            return new_sample, y[random_index[0]]
+
+        x_temp, y_temp = upsample_core(x, y, min_samples, applySmote, size=1)
+
+    # Append the temp list to the original data
+    if len(x_temp) > 0:
+        x = np.vstack((x, np.array(x_temp)))
+        y = np.vstack((y, np.array(y_temp)))
+
+    # Shuffle data
+    indices = np.arange(len(x))
+    rng.shuffle(indices)
+    x = x[indices]
+    y = y[indices]
+
+    del x_temp
+    del y_temp
+
+    return x, y
+
+
+def save_model_params(path):
+    """Saves the model parameters to a file.
+
+    Args:
+        path: Path to the file.
+    """
+    utils.save_params(
+        path,
+        (
+            "Hidden units",
+            "Dropout",
+            "Batchsize",
+            "Learning rate",
+            "Crop mode",
+            "Crop overlap",
+            "Audio speed",
+            "Upsamling mode",
+            "Upsamling ratio",
+            "use mixup",
+            "use label smoothing",
+            "BirdNET Model version",
+        ),
+        (
+            cfg.TRAIN_HIDDEN_UNITS,
+            cfg.TRAIN_DROPOUT,
+            cfg.TRAIN_BATCH_SIZE,
+            cfg.TRAIN_LEARNING_RATE,
+            cfg.SAMPLE_CROP_MODE,
+            cfg.SIG_OVERLAP,
+            cfg.AUDIO_SPEED,
+            cfg.UPSAMPLING_MODE,
+            cfg.UPSAMPLING_RATIO,
+            cfg.TRAIN_WITH_MIXUP,
+            cfg.TRAIN_WITH_LABEL_SMOOTHING,
+            cfg.MODEL_VERSION,
+        ),
+    )
+
+
+def reset_custom_classifier():
+    """
+    Resets the custom classifier by setting the global variables C_INTERPRETER and C_PBMODEL to None.
+    This function is used to clear any existing custom classifier models and interpreters, effectively
+    resetting the state of the custom classifier.
+    """
+    global C_INTERPRETER
+    global C_PBMODEL
+
+    C_INTERPRETER = None
+    C_PBMODEL = None
+
+
+def load_model(class_output=True):
+    """
+    Loads the machine learning model based on the configuration provided.
+    This function loads either a TensorFlow Lite (TFLite) model or a protobuf model
+    depending on the file extension of the model path specified in the configuration.
+    It sets up the global variables for the model interpreter and input/output layer indices.
+
+    Args:
+        class_output (bool): If True, sets the output layer index to the classification output.
+                             If False, sets the output layer index to the feature embeddings.
+    """
+    global PBMODEL
+    global INTERPRETER
+    global INPUT_LAYER_INDEX
+    global OUTPUT_LAYER_INDEX
+    global OUTPUT_DETAILS
+
+    # Do we have to load the tflite or protobuf model?
+    if cfg.MODEL_PATH.endswith(".tflite"):
+        if not INTERPRETER:
+            # Load TFLite model and allocate tensors.
+            INTERPRETER = tflite.Interpreter(
+                model_path=os.path.join(SCRIPT_DIR, cfg.MODEL_PATH), num_threads=cfg.TFLITE_THREADS
+            )
+            INTERPRETER.allocate_tensors()
+
+            # Get input and output tensors.
+            input_details = INTERPRETER.get_input_details()
+            OUTPUT_DETAILS = INTERPRETER.get_output_details()
+
+            # Get input tensor index
+            INPUT_LAYER_INDEX = input_details[0]["index"]
+
+        # Get classification output or feature embeddings
+        OUTPUT_LAYER_INDEX = OUTPUT_DETAILS[0]["index"] if class_output else OUTPUT_DETAILS[0]["index"] - 1
+
+    elif not PBMODEL:
+        # Load protobuf model
+        # Note: This will throw a bunch of warnings about custom gradients
+        # which we will ignore until TF lets us block them
+        PBMODEL = keras.models.load_model(os.path.join(SCRIPT_DIR, cfg.MODEL_PATH), compile=False)
+
+
+def load_custom_classifier():
+    """
+    Loads a custom classifier model based on the file extension of the provided model path.
+    If the model file ends with ".tflite", it loads a TensorFlow Lite model and sets up the interpreter,
+    input layer index, output layer index, and input size.
+    If the model file does not end with ".tflite", it loads a TensorFlow SavedModel.
+    """
+    global C_INTERPRETER
+    global C_INPUT_LAYER_INDEX
+    global C_OUTPUT_LAYER_INDEX
+    global C_INPUT_SIZE
+    global C_PBMODEL
+
+    if cfg.CUSTOM_CLASSIFIER.endswith(".tflite"):
+        # Load TFLite model and allocate tensors.
+        C_INTERPRETER = tflite.Interpreter(model_path=cfg.CUSTOM_CLASSIFIER, num_threads=cfg.TFLITE_THREADS)
+        C_INTERPRETER.allocate_tensors()
+
+        # Get input and output tensors.
+        input_details = C_INTERPRETER.get_input_details()
+        output_details = C_INTERPRETER.get_output_details()
+
+        # Get input tensor index
+        C_INPUT_LAYER_INDEX = input_details[0]["index"]
+
+        C_INPUT_SIZE = input_details[0]["shape"][-1]
+
+        # Get classification output
+        C_OUTPUT_LAYER_INDEX = output_details[0]["index"]
+    else:
+        import tensorflow as tf
+
+        tf.get_logger().setLevel("ERROR")
+
+        C_PBMODEL = tf.saved_model.load(cfg.CUSTOM_CLASSIFIER)
+
+
+def load_meta_model():
+    """Loads the model for species prediction.
+
+    Initializes the model used to predict species list, based on coordinates and week of year.
+    """
+    global M_INTERPRETER
+    global M_INPUT_LAYER_INDEX
+    global M_OUTPUT_LAYER_INDEX
+
+    # Load TFLite model and allocate tensors.
+    M_INTERPRETER = tflite.Interpreter(
+        model_path=os.path.join(SCRIPT_DIR, cfg.MDATA_MODEL_PATH), num_threads=cfg.TFLITE_THREADS
+    )
+    M_INTERPRETER.allocate_tensors()
+
+    # Get input and output tensors.
+    input_details = M_INTERPRETER.get_input_details()
+    output_details = M_INTERPRETER.get_output_details()
+
+    # Get input tensor index
+    M_INPUT_LAYER_INDEX = input_details[0]["index"]
+    M_OUTPUT_LAYER_INDEX = output_details[0]["index"]
+
+
+def build_linear_classifier(num_labels, input_size, hidden_units=0, dropout=0.0):
+    """Builds a classifier.
+
+    Args:
+        num_labels: Output size.
+        input_size: Size of the input.
+        hidden_units: If > 0, creates another hidden layer with the given number of units.
+        dropout: Dropout rate.
+
+    Returns:
+        A new classifier.
+    """
+    # import keras
+    from tensorflow import keras
+
+    # Build a simple one- or two-layer linear classifier
+    model = keras.Sequential()
+
+    # Input layer
+    model.add(keras.layers.InputLayer(input_shape=(input_size,)))
+
+    # Batch normalization on input to standardize embeddings
+    model.add(keras.layers.BatchNormalization())
+
+    # Optional L2 regularization for all dense layers
+    regularizer = keras.regularizers.l2(1e-5)
+
+    # Hidden layer with improved architecture
+    if hidden_units > 0:
+        # Dropout layer before hidden layer
+        if dropout > 0:
+            model.add(keras.layers.Dropout(dropout))
+
+        # Add a hidden layer with L2 regularization
+        model.add(
+            keras.layers.Dense(
+                hidden_units, activation="relu", kernel_regularizer=regularizer, kernel_initializer="he_normal"
+            )
+        )
+
+        # Add another batch normalization after the hidden layer
+        model.add(keras.layers.BatchNormalization())
+
+    # Dropout layer before output
+    if dropout > 0:
+        model.add(keras.layers.Dropout(dropout))
+
+    # Classification layer with L2 regularization
+    model.add(keras.layers.Dense(num_labels, kernel_regularizer=regularizer, kernel_initializer="glorot_uniform"))
+
+    # Activation layer
+    model.add(keras.layers.Activation("sigmoid"))
+
+    return model
+
+
+def train_linear_classifier(
+    classifier,
+    x_train,
+    y_train,
+    x_test,
+    y_test,
+    epochs,
+    batch_size,
+    learning_rate,
+    val_split,
+    upsampling_ratio,
+    upsampling_mode,
+    train_with_mixup,
+    train_with_label_smoothing,
+    train_with_focal_loss=False,
+    focal_loss_gamma=2.0,
+    focal_loss_alpha=0.25,
+    on_epoch_end=None,
+):
+    """Trains a custom classifier.
+
+    Trains a new classifier for BirdNET based on the given data.
+
+    Args:
+        classifier: The classifier to be trained.
+        x_train: Samples.
+        y_train: Labels.
+        x_test: Validation samples.
+        y_test: Validation labels.
+        epochs: Number of epochs to train.
+        batch_size: Batch size.
+        learning_rate: The learning rate during training.
+        val_split: Validation split ratio (is 0 when using test data).
+        upsampling_ratio: Upsampling ratio.
+        upsampling_mode: Upsampling mode.
+        train_with_mixup: If True, applies mixup to the training data.
+        train_with_label_smoothing: If True, applies label smoothing to the training data.
+        train_with_focal_loss: If True, uses focal loss instead of binary cross-entropy loss.
+        focal_loss_gamma: Focal loss gamma parameter.
+        focal_loss_alpha: Focal loss alpha parameter.
+        on_epoch_end: Optional callback `function(epoch, logs)`.
+
+    Returns:
+        (classifier, history)
+    """
+    # import keras
+    from tensorflow import keras
+
+    class FunctionCallback(keras.callbacks.Callback):
+        def __init__(self, on_epoch_end=None) -> None:
+            super().__init__()
+            self.on_epoch_end_fn = on_epoch_end
+
+        def on_epoch_end(self, epoch, logs=None):
+            if self.on_epoch_end_fn:
+                self.on_epoch_end_fn(epoch, logs)
+
+    # Set random seed
+    rng = np.random.default_rng(cfg.RANDOM_SEED)
+
+    # Shuffle data
+    idx = np.arange(x_train.shape[0])
+    rng.shuffle(idx)
+    x_train = x_train[idx]
+    y_train = y_train[idx]
+
+    # Random val split
+    if val_split > 0:
+        if not cfg.MULTI_LABEL:
+            x_train, y_train, x_val, y_val = random_split(x_train, y_train, val_split)
+        else:
+            x_train, y_train, x_val, y_val = random_multilabel_split(x_train, y_train, val_split)
+    else:
+        x_val = x_test
+        y_val = y_test
+
+    print(
+        f"Training on {x_train.shape[0]} samples, validating on {x_val.shape[0]} samples.",
+        flush=True,
+    )
+
+    # Upsample training data
+    if upsampling_ratio > 0:
+        x_train, y_train = upsampling(x_train, y_train, upsampling_ratio, upsampling_mode)
+        print(f"Upsampled training data to {x_train.shape[0]} samples.", flush=True)
+
+    # Apply mixup to training data
+    if train_with_mixup and not cfg.BINARY_CLASSIFICATION:
+        x_train, y_train = mixup(x_train, y_train)
+
+    # Apply label smoothing
+    if train_with_label_smoothing and not cfg.BINARY_CLASSIFICATION:
+        y_train = label_smoothing(y_train)
+
+    # Early stopping with patience depending on dataset size
+    patience = min(10, max(5, int(epochs / 10)))
+    min_delta = 0.001
+
+    callbacks = [
+        # EarlyStopping with restore_best_weights
+        keras.callbacks.EarlyStopping(
+            monitor="val_AUPRC",
+            mode="max",
+            patience=patience,
+            verbose=1,
+            min_delta=min_delta,
+            restore_best_weights=True,
+        ),
+        # Function callback for progress tracking
+        FunctionCallback(on_epoch_end=on_epoch_end),
+    ]
+
+    # Learning rate schedule - use cosine decay with warmup
+    warmup_epochs = min(5, int(epochs * 0.1))
+
+    def lr_schedule(epoch, lr):
+        if epoch < warmup_epochs:
+            # Linear warmup
+            return learning_rate * (epoch + 1) / warmup_epochs
+
+        # Cosine decay
+        progress = (epoch - warmup_epochs) / (epochs - warmup_epochs)
+        return learning_rate * (0.1 + 0.9 * (1 + np.cos(np.pi * progress)) / 2)
+
+    # Add LR scheduler callback
+    callbacks.append(keras.callbacks.LearningRateScheduler(lr_schedule))
+
+    optimizer_cls = keras.optimizers.legacy.Adam if sys.platform == "darwin" else keras.optimizers.Adam
+
+    def _focal_loss(y_true, y_pred):
+        return focal_loss(y_true, y_pred, gamma=cfg.FOCAL_LOSS_GAMMA, alpha=cfg.FOCAL_LOSS_ALPHA)
+
+    # Choose the loss function based on config
+    loss_function = _focal_loss if train_with_focal_loss else custom_loss
+
+    # Compile model with appropriate metrics for classification task
+    classifier.compile(
+        optimizer=optimizer_cls(learning_rate=learning_rate),
+        loss=loss_function,
+        metrics=[
+            keras.metrics.AUC(
+                curve="PR",
+                multi_label=cfg.MULTI_LABEL,
+                name="AUPRC",
+                num_labels=y_train.shape[1] if cfg.MULTI_LABEL else None,
+                from_logits=True,
+            ),
+            keras.metrics.AUC(
+                curve="ROC",
+                multi_label=cfg.MULTI_LABEL,
+                name="AUROC",
+                num_labels=y_train.shape[1] if cfg.MULTI_LABEL else None,
+                from_logits=True,
+            ),
+        ],
+    )
+
+    # Train model
+    history = classifier.fit(
+        x_train, y_train, epochs=epochs, batch_size=batch_size, validation_data=(x_val, y_val), callbacks=callbacks
+    )
+
+    return classifier, history
+
+
+def save_linear_classifier(classifier, model_path: str, labels: list[str], mode="replace"):
+    """Saves the classifier as a tflite model, as well as the used labels in a .txt.
+
+    Args:
+        classifier: The custom classifier.
+        model_path: Path the model will be saved at.
+        labels: List of labels used for the classifier.
+    """
+    import tensorflow as tf
+
+    global PBMODEL
+
+    tf.get_logger().setLevel("ERROR")
+
+    if PBMODEL is None:
+        PBMODEL = tf.keras.models.load_model(os.path.join(SCRIPT_DIR, cfg.PB_MODEL), compile=False)
+
+    saved_model = PBMODEL
+
+    # Remove activation layer
+    classifier.pop()
+
+    if mode == "replace":
+        combined_model = tf.keras.Sequential([saved_model.embeddings_model, classifier], "basic")
+    elif mode == "append":
+        intermediate = classifier(saved_model.model.get_layer("GLOBAL_AVG_POOL").output)
+
+        output = tf.keras.layers.concatenate([saved_model.model.output, intermediate], name="combined_output")
+
+        combined_model = tf.keras.Model(inputs=saved_model.model.input, outputs=output)
+    else:
+        raise ValueError("Model save mode must be either 'replace' or 'append'")
+
+    # Append .tflite if necessary
+    if not model_path.endswith(".tflite"):
+        model_path += ".tflite"
+
+    # Make folders
+    os.makedirs(os.path.dirname(model_path), exist_ok=True)
+
+    # Save model as tflite
+    converter = tf.lite.TFLiteConverter.from_keras_model(combined_model)
+    tflite_model = converter.convert()
+
+    with open(model_path, "wb") as f:
+        f.write(tflite_model)
+
+    if mode == "append":
+        labels = [*utils.read_lines(os.path.join(SCRIPT_DIR, cfg.LABELS_FILE)), *labels]
+
+    # Save labels
+    with open(model_path.replace(".tflite", "_Labels.txt"), "w", encoding="utf-8") as f:
+        f.writelines(label + "\n" for label in labels)
+
+    save_model_params(model_path.replace(".tflite", "_Params.csv"))
+
+
+def save_raven_model(classifier, model_path: str, labels: list[str], mode="replace"):
+    """
+    Save a TensorFlow model with a custom classifier and associated metadata for use with BirdNET.
+
+    Args:
+        classifier (tf.keras.Model): The custom classifier model to be saved.
+        model_path (str): The path where the model will be saved.
+        labels (list[str]): A list of labels associated with the classifier.
+        mode (str, optional): The mode for saving the model. Can be either "replace" or "append".
+                              Defaults to "replace".
+
+    Raises:
+        ValueError: If the mode is not "replace" or "append".
+
+    Returns:
+        None
+    """
+    import csv
+    import json
+
+    import tensorflow as tf
+
+    global PBMODEL
+
+    tf.get_logger().setLevel("ERROR")
+
+    if PBMODEL is None:
+        PBMODEL = tf.keras.models.load_model(os.path.join(SCRIPT_DIR, cfg.PB_MODEL), compile=False)
+
+    saved_model = PBMODEL
+
+    if mode == "replace":
+        combined_model = tf.keras.Sequential([saved_model.embeddings_model, classifier], "basic")
+    elif mode == "append":
+        # Remove activation layer
+        classifier.pop()
+        intermediate = classifier(saved_model.model.get_layer("GLOBAL_AVG_POOL").output)
+
+        output = tf.keras.layers.concatenate([saved_model.model.output, intermediate], name="combined_output")
+
+        combined_model = tf.keras.Model(inputs=saved_model.model.input, outputs=output)
+    else:
+        raise ValueError("Model save mode must be either 'replace' or 'append'")
+
+    # Make signatures
+    class SignatureModule(tf.Module):
+        def __init__(self, keras_model):
+            super().__init__()
+            self.model = keras_model
+
+        @tf.function(input_signature=[tf.TensorSpec(shape=[None, 144000], dtype=tf.float32)])
+        def basic(self, inputs):
+            return {"scores": self.model(inputs)}
+
+    smodel = SignatureModule(combined_model)
+    signatures = {
+        "basic": smodel.basic,
+    }
+
+    # Save signature model
+    os.makedirs(os.path.dirname(model_path), exist_ok=True)
+    model_path = model_path.removesuffix(".tflite")
+    tf.saved_model.save(smodel, model_path, signatures=signatures)
+
+    if mode == "append":
+        labels = [*utils.read_lines(os.path.join(SCRIPT_DIR, cfg.LABELS_FILE)), *labels]
+
+    # Save label file
+    labelIds = [label[:4].replace(" ", "") + str(i) for i, label in enumerate(labels, 1)]
+    labels_dir = os.path.join(model_path, "labels")
+
+    os.makedirs(labels_dir, exist_ok=True)
+
+    with open(os.path.join(labels_dir, "label_names.csv"), "w", newline="") as labelsfile:
+        labelwriter = csv.writer(labelsfile)
+        labelwriter.writerows(zip(labelIds, labels, strict=True))
+
+    # Save class names file
+    classes_dir = os.path.join(model_path, "classes")
+
+    os.makedirs(classes_dir, exist_ok=True)
+
+    with open(os.path.join(classes_dir, "classes.csv"), "w", newline="") as classesfile:
+        classeswriter = csv.writer(classesfile)
+        for labelId in labelIds:
+            classeswriter.writerow((labelId, 0.25, cfg.SIG_FMIN, cfg.SIG_FMAX, False))
+
+    # Save model config
+    model_config = os.path.join(model_path, "model_config.json")
+
+    with open(model_config, "w") as modelconfigfile:
+        modelconfig = {
+            "specVersion": 1,
+            "modelDescription": "Custom classifier trained with BirdNET "
+            + cfg.MODEL_VERSION
+            + " embeddings.\n"
+            + "BirdNET was developed by the K. Lisa Yang Center for Conservation Bioacoustics"
+            + "at the Cornell Lab of Ornithology in collaboration with Chemnitz University of Technology.\n\n"
+            + "https://birdnet.cornell.edu",
+            "modelTypeConfig": {"modelType": "RECOGNITION"},
+            "signatures": [
+                {
+                    "signatureName": "basic",
+                    "modelInputs": [
+                        {
+                            "inputName": "inputs",
+                            "sampleRate": 48000.0,
+                            "inputConfig": ["batch", "samples"],
+                        }
+                    ],
+                    "modelOutputs": [{"outputName": "scores", "outputType": "SCORES"}],
+                }
+            ],
+            "globalSemanticKeys": labelIds,
+        }
+        json.dump(modelconfig, modelconfigfile, indent=2)
+
+        model_params = os.path.join(model_path, "model_params.csv")
+
+        save_model_params(model_params)
+
+
+def predict_filter(lat, lon, week):
+    """Predicts the probability for each species.
+
+    Args:
+        lat: The latitude.
+        lon: The longitude.
+        week: The week of the year [1-48]. Use -1 for yearlong.
+
+    Returns:
+        A list of probabilities for all species.
+    """
+    # Does interpreter exist?
+    if M_INTERPRETER is None:
+        load_meta_model()
+
+    # Prepare mdata as sample
+    sample = np.expand_dims(np.array([lat, lon, week], dtype="float32"), 0)
+
+    # Run inference
+    M_INTERPRETER.set_tensor(M_INPUT_LAYER_INDEX, sample)
+    M_INTERPRETER.invoke()
+
+    return M_INTERPRETER.get_tensor(M_OUTPUT_LAYER_INDEX)[0]
+
+
+def explore(lat: float, lon: float, week: int):
+    """Predicts the species list.
+
+    Predicts the species list based on the coordinates and week of year.
+
+    Args:
+        lat: The latitude.
+        lon: The longitude.
+        week: The week of the year [1-48]. Use -1 for yearlong.
+
+    Returns:
+        A sorted list of tuples with the score and the species.
+    """
+    # Make filter prediction
+    l_filter = predict_filter(lat, lon, week)
+
+    # Apply threshold
+    l_filter = np.where(l_filter >= cfg.LOCATION_FILTER_THRESHOLD, l_filter, 0)
+
+    # Zip with labels
+    l_filter = list(zip(l_filter, cfg.LABELS, strict=True))
+
+    # Sort by filter value
+    return sorted(l_filter, key=lambda x: x[0], reverse=True)
+
+
+def focal_loss(y_true, y_pred, gamma=2.0, alpha=0.25, epsilon=1e-7):
+    """
+    Focal loss for better handling of class imbalance.
+
+    This loss function gives more weight to hard examples and down-weights easy examples.
+    Particularly helpful for imbalanced datasets where some classes have few samples.
+
+    Args:
+        y_true: Ground truth labels.
+        y_pred: Predicted probabilities.
+        gamma: Focusing parameter. Higher values mean more focus on hard examples.
+        alpha: Balance parameter. Controls weight of positive vs negative examples.
+        epsilon: Small constant to prevent log(0).
+
+    Returns:
+        Focal loss value.
+    """
+    import tensorflow.keras.backend as K
+
+    # Apply sigmoid if not already applied
+    y_pred = K.clip(y_pred, epsilon, 1.0 - epsilon)
+
+    # Calculate cross entropy
+    cross_entropy = -y_true * K.log(y_pred) - (1 - y_true) * K.log(1 - y_pred)
+
+    # Calculate focal weight
+    p_t = y_true * y_pred + (1 - y_true) * (1 - y_pred)
+    focal_weight = K.pow(1 - p_t, gamma)
+
+    # Apply alpha balancing
+    alpha_factor = y_true * alpha + (1 - y_true) * (1 - alpha)
+
+    # Calculate focal loss
+    focal_loss = alpha_factor * focal_weight * cross_entropy
+
+    # Sum over all classes
+    return K.sum(focal_loss, axis=-1)
+
+
+def custom_loss(y_true, y_pred, epsilon=1e-7):
+    import tensorflow.keras.backend as K
+
+    # Calculate loss for positive labels with epsilon
+    positive_loss = -K.sum(y_true * K.log(K.clip(y_pred, epsilon, 1.0 - epsilon)), axis=-1)
+
+    # Calculate loss for negative labels with epsilon
+    negative_loss = -K.sum((1 - y_true) * K.log(K.clip(1 - y_pred, epsilon, 1.0 - epsilon)), axis=-1)
+
+    # Combine both loss terms
+    return positive_loss + negative_loss
+
+
+def flat_sigmoid(x, sensitivity=-1, bias=1.0):
+    """
+    Applies a flat sigmoid function to the input array with a bias shift.
+
+    The flat sigmoid function is defined as:
+        f(x) = 1 / (1 + exp(sensitivity * clip(x + bias, -20, 20)))
+
+    We transform the bias parameter to a range of [-100, 100] with the formula:
+        transformed_bias = (bias - 1.0) * 10.0
+
+    Thus, higher bias values will shift the sigmoid function to the right on the x-axis, making it more "sensitive".
+
+    Note: Not sure why we are clipping, must be for numerical stability somewhere else in the code.
+
+    Args:
+        x (array-like): Input data.
+        sensitivity (float, optional): Sensitivity parameter for the sigmoid function. Default is -1.
+        bias (float, optional): Bias parameter to shift the sigmoid function on the x-axis. Must be in the range [0.01, 1.99]. Default is 1.0.
+
+    Returns:
+        numpy.ndarray: Transformed data after applying the flat sigmoid function.
+    """
+
+    transformed_bias = (bias - 1.0) * 10.0
+
+    return 1 / (1.0 + np.exp(sensitivity * np.clip(x + transformed_bias, -20, 20)))
+
+
+def predict(sample):
+    """Uses the main net to predict a sample.
+
+    Args:
+        sample: Audio sample.
+
+    Returns:
+        The prediction scores for the sample.
+    """
+    # Has custom classifier?
+    if cfg.CUSTOM_CLASSIFIER is not None:
+        return predict_with_custom_classifier(sample)
+
+    load_model()
+
+    if PBMODEL is None:
+        # Reshape input tensor
+        INTERPRETER.resize_tensor_input(INPUT_LAYER_INDEX, [len(sample), *sample[0].shape])
+        INTERPRETER.allocate_tensors()
+
+        # Make a prediction (Audio only for now)
+        INTERPRETER.set_tensor(INPUT_LAYER_INDEX, np.array(sample, dtype="float32"))
+        INTERPRETER.invoke()
+        return INTERPRETER.get_tensor(OUTPUT_LAYER_INDEX)
+
+    # Make a prediction (Audio only for now)
+    return PBMODEL.basic(sample)["scores"]
+
+
+def predict_with_custom_classifier(sample):
+    """Uses the custom classifier to make a prediction.
+
+    Args:
+        sample: Audio sample.
+
+    Returns:
+        The prediction scores for the sample.
+    """
+    # Does interpreter exist?
+    if C_INTERPRETER is None and C_PBMODEL is None:
+        load_custom_classifier()
+
+    if C_PBMODEL is None:
+        vector = embeddings(sample) if C_INPUT_SIZE != 144000 else sample
+
+        # Reshape input tensor
+        C_INTERPRETER.resize_tensor_input(C_INPUT_LAYER_INDEX, [len(vector), *vector[0].shape])
+        C_INTERPRETER.allocate_tensors()
+
+        # Make a prediction
+        C_INTERPRETER.set_tensor(C_INPUT_LAYER_INDEX, np.array(vector, dtype="float32"))
+        C_INTERPRETER.invoke()
+
+        return C_INTERPRETER.get_tensor(C_OUTPUT_LAYER_INDEX)
+
+    return C_PBMODEL.basic(sample)["scores"]
+
+
+def embeddings(sample):
+    """Extracts the embeddings for a sample.
+
+    Args:
+        sample: Audio samples.
+
+    Returns:
+        The embeddings.
+    """
+
+    load_model(False)
+
+    sample = np.array(sample, dtype="float32")
+
+    # Reshape input tensor
+    INTERPRETER.resize_tensor_input(INPUT_LAYER_INDEX, [len(sample), *sample[0].shape])
+    INTERPRETER.allocate_tensors()
+
+    # Extract feature embeddings
+    INTERPRETER.set_tensor(INPUT_LAYER_INDEX, sample)
+    INTERPRETER.invoke()
+
+    return INTERPRETER.get_tensor(OUTPUT_LAYER_INDEX)