bnn_models_built_in_utils

Description of helper functions for Bayesian modeling implemented in the module bnn_for_14C_calibration.bnn_models_built_in_utils:

bnn_for_14C_calibration.bnn_models_built_in_utils

bnn_load_predictions_(filepath)

Load predictions generated by a Bayesian neural network from a CSV file.

Parameters:

Name	Type	Description	Default
filepath	str	Path to the CSV file containing predictions.	required

Returns:

Name	Type	Description
predictions_array	ndarray	Predictions as a float32 numpy array.
nb_intervals	int	Number of intervals (rows) in the predictions.
nb_curves	int	Number of curves (columns) in the predictions.

Notes

• Intended for predictions generated by bnn_make_predictions_ for calibration purposes.

Source code in src/bnn_for_14C_calibration/bnn_models_built_in_utils.py

def bnn_load_predictions_(filepath: str) -> Tuple[np.ndarray, int, int]:
    """
    Load predictions generated by a Bayesian neural network from a CSV file.

    Parameters
    ----------
    filepath : str
        Path to the CSV file containing predictions.

    Returns
    -------
    predictions_array : np.ndarray
        Predictions as a float32 numpy array.
    nb_intervals : int
        Number of intervals (rows) in the predictions.
    nb_curves : int
        Number of curves (columns) in the predictions.

    Notes
    -----
    - Intended for predictions generated by `bnn_make_predictions_` for calibration purposes.
    """
    predictions_df = load_data(path=filepath, sep=",")
    predictions_array = predictions_df.to_numpy(dtype=np.float32)
    nb_intervals, nb_curves = predictions_array.shape
    return predictions_array, nb_intervals, nb_curves

bnn_make_predictions_(bnn_model, X_test, iterations=100, batch_size=None)

Generate predictions from a Bayesian neural network by repeated stochastic forward passes.

Parameters:

Name	Type	Description	Default
bnn_model	Model	Trained Bayesian neural network.	required
X_test	ndarray or Tensor	Test input data.	required
iterations	int	Number of stochastic forward passes to perform (default 100).	100
batch_size	int	Batch size to use during prediction. If None, defaults to full batch.	None

Returns:

Type	Description
ndarray	Concatenated predictions from all iterations. Shape: (n_samples, n_outputs, iterations)

Notes

• Each call to the model produces a stochastic output due to the variational posterior.

Source code in src/bnn_for_14C_calibration/bnn_models_built_in_utils.py

def bnn_make_predictions_(
    bnn_model: keras.Model,
    X_test: Union[np.ndarray, tf.Tensor],
    iterations: int = 100,
    batch_size: int = None
) -> np.ndarray:
    """
    Generate predictions from a Bayesian neural network by repeated stochastic forward passes.

    Parameters
    ----------
    bnn_model : keras.Model
        Trained Bayesian neural network.
    X_test : np.ndarray or tf.Tensor
        Test input data.
    iterations : int, optional
        Number of stochastic forward passes to perform (default 100).
    batch_size : int, optional
        Batch size to use during prediction. If None, defaults to full batch.

    Returns
    -------
    np.ndarray
        Concatenated predictions from all iterations. Shape: (n_samples, n_outputs, iterations)

    Notes
    -----
    - Each call to the model produces a stochastic output due to the variational posterior.
    """
    predicted = []
    for _ in range(iterations):
        # !!! TO DO : investiger les différences de comportement entre 
        # model et model.predict avec batch_size != None !!!
        # predicted.append(bnn_model.predict(X_test, batch_size=batch_size, verbose=0))
        predicted.append(bnn_model(X_test))
    predicted = np.concatenate(predicted, axis=1)

    return predicted

gaussian_prior(kernel_size, bias_size, dtype=None, sigma=1.0)

Define a Gaussian prior distribution over weights and biases of a Bayesian neural network.

Parameters:

Name	Type	Description	Default
kernel_size	int	Number of kernel weights in the layer.	required
bias_size	int	Number of bias terms in the layer.	required
dtype	DType	TensorFlow data type for the distribution (default None).	None
sigma	float	Standard deviation of the Gaussian prior (default 1.0).	1.0

Returns:

Type	Description
Sequential	Non-trainable prior distribution model. Each weight and bias is assumed independent and distributed according to a normal distribution N(0, sigma^2).

Notes

• The prior distribution is non-trainable; its parameters are fixed.

Source code in src/bnn_for_14C_calibration/bnn_models_built_in_utils.py

def gaussian_prior(
    kernel_size: int,
    bias_size: int,
    dtype: tf.dtypes.DType = None,
    sigma: float = 1.0
) -> keras.Sequential:
    """
    Define a Gaussian prior distribution over weights and biases of a Bayesian neural network.

    Parameters
    ----------
    kernel_size : int
        Number of kernel weights in the layer.
    bias_size : int
        Number of bias terms in the layer.
    dtype : tf.dtypes.DType, optional
        TensorFlow data type for the distribution (default None).
    sigma : float, optional
        Standard deviation of the Gaussian prior (default 1.0).

    Returns
    -------
    keras.Sequential
        Non-trainable prior distribution model. Each weight and bias is assumed independent
        and distributed according to a normal distribution N(0, sigma^2).

    Notes
    -----
    - The prior distribution is **non-trainable**; its parameters are fixed.
    """
    n = kernel_size + bias_size
    var_mat_diag = sigma * tf.ones(n, dtype=dtype)
    prior_model = keras.Sequential(
        [
            tfp.layers.DistributionLambda(
                lambda t: tfp.distributions.MultivariateNormalDiag(
                    loc=tf.zeros(n, dtype=dtype), scale_diag=var_mat_diag
                )
            )
        ]
    )
    return prior_model

independent_gaussian_posterior(kernel_size, bias_size, dtype=None)

Define an independent Gaussian posterior distribution for variational inference in Bayesian neural networks.

Parameters:

Name	Type	Description	Default
kernel_size	int	Number of kernel weights in the layer.	required
bias_size	int	Number of bias terms in the layer.	required
dtype	DType	TensorFlow data type for the distribution (default None).	None

Returns:

Type	Description
Sequential	Trainable posterior distribution model. Each weight and bias has a learnable mean and diagonal variance. Off-diagonal covariances are zero, implying independence among weights.

Notes

• The parameters of the distribution (mean and diagonal variance) are trainable.

Source code in src/bnn_for_14C_calibration/bnn_models_built_in_utils.py

def independent_gaussian_posterior(
    kernel_size: int,
    bias_size: int,
    dtype: tf.dtypes.DType = None
) -> keras.Sequential:
    """
    Define an independent Gaussian posterior distribution for variational inference
    in Bayesian neural networks.

    Parameters
    ----------
    kernel_size : int
        Number of kernel weights in the layer.
    bias_size : int
        Number of bias terms in the layer.
    dtype : tf.dtypes.DType, optional
        TensorFlow data type for the distribution (default None).

    Returns
    -------
    keras.Sequential
        Trainable posterior distribution model. Each weight and bias has a learnable mean
        and diagonal variance. Off-diagonal covariances are zero, implying independence
        among weights.

    Notes
    -----
    - The parameters of the distribution (mean and diagonal variance) are **trainable**.
    """
    n = kernel_size + bias_size
    posterior_model = keras.Sequential(
        [
            tfp.layers.VariableLayer(
                tfp.layers.IndependentNormal.params_size(n), dtype=dtype
            ),
            tfp.layers.IndependentNormal(n),
        ]
    )
    return posterior_model

negative_loglikelihood(targets, estimated_distribution)

Compute negative log-likelihood for a probabilistic prediction.

Parameters:

Name	Type	Description	Default
targets	ndarray or Tensor	True target values.	required
estimated_distribution	Distribution	Estimated (predicted) distribution output from the Bayesian network.	required

Returns:

Type	Description
Tensor	Negative log-likelihood value for use as a loss function in Bayesian neural networks.

Notes

• Used with stochastic outputs of probabilistic networks.
• Supports variational inference for posterior estimation.

Source code in src/bnn_for_14C_calibration/bnn_models_built_in_utils.py

def negative_loglikelihood(
    targets: Union[np.ndarray, tf.Tensor],
    estimated_distribution: tfp.distributions.Distribution
) -> tf.Tensor:
    """
    Compute negative log-likelihood for a probabilistic prediction.

    Parameters
    ----------
    targets : np.ndarray or tf.Tensor
        True target values.
    estimated_distribution : tfp.distributions.Distribution
        Estimated (predicted) distribution output from the Bayesian network.

    Returns
    -------
    tf.Tensor
        Negative log-likelihood value for use as a loss function in Bayesian neural networks.

    Notes
    -----
    - Used with stochastic outputs of probabilistic networks.
    - Supports variational inference for posterior estimation.
    """
    return -estimated_distribution.log_prob(targets)