Internal training routine for the baseline SPQR (Spline-based Probabilistic Quantile Regression) model. This function builds the neural network architecture, compiles it with the SPQR negative log-likelihood loss, and performs supervised training using precomputed spline basis matrices.
Usage
in.fit.spqr(
input_dim,
hidden_dim,
n.knots,
knots,
x_training,
x_validation,
y_training,
y_validation,
m_basis_training,
m_basis_validation,
i_basis_training,
i_basis_validation,
hyperparameter = NULL
)Arguments
- input_dim
Integer. Number of covariates (input features).
Integer vector. Number of units in each hidden layer.
- n.knots
Integer. Total number of spline basis functions.
- knots
Numeric vector of spline knot locations.
- x_training
Numeric matrix of training covariates.
- x_validation
Numeric matrix of validation covariates.
- y_training
Numeric vector or matrix of training responses.
- y_validation
Numeric vector or matrix of validation responses.
- m_basis_training
Numeric matrix of spline basis evaluations for the training responses.
- m_basis_validation
Numeric matrix of spline basis evaluations for the validation responses.
- i_basis_training
Numeric matrix of integrated spline basis evaluations for the training responses.
- i_basis_validation
Numeric matrix of integrated spline basis evaluations for the validation responses.
- hyperparameter
List. Model training configuration created by
create.packages.hyperparameter(), containing elements such as `epochs`, `batch_size`, and `activation`.
Details
This function assumes that:
Covariates and responses have already been preprocessed.
Spline basis matrices (`m_basis_*`, `i_basis_*`) have already been constructed.
Hyperparameters are supplied via a structured list.
It is not intended to be called directly by end users.
The model is trained using the nloglik_loss_SPQR loss
function and the Adam optimizer. Early stopping is applied
based on validation loss, and the best-performing weights are
saved during training.