Performs train / calibration / test split, fits the model with early stopping and learning-rate reduction, computes test metrics and conformal quantiles, and saves the model + scaler + metrics to disk.
Usage
train_soilVAE(
X,
y,
family_id,
prop,
out_dir = "models",
epochs = 30L,
batch_size = 32L,
val_split = 0.2,
patience_es = 8L,
patience_rl = 4L,
latent_dim = 16L,
min_n = 30L
)Arguments
- X
Numeric matrix (rows = samples, cols = preprocessed bands)
- y
Numeric vector of soil property values (length == nrow(X))
- family_id
Family identifier string (used for output directory)
- prop
Soil property name (used for output file names)
- out_dir
Root output directory (default: "models")
- epochs
Max training epochs (default 30)
- batch_size
Mini-batch size (default 32)
- val_split
Validation fraction from training data (default 0.2)
- patience_es
Early stopping patience (default 8)
- patience_rl
LR reduction patience (default 4)
- latent_dim
Latent dimension (default 16)
- min_n
Minimum samples required to attempt training (default 30)