soilVAE 0.2.0

This release fixes defects that made 0.1.x predict poorly with its default settings, and removes the bundled dataset: the package now contains code only.

On prospectr::NIRsoil (soil carbon, SNV spectra), the default settings now give a median test R² of 0.78 (0.69-0.82 over 10 seeds) on the calibration/test split shipped with the data, against 0.68 for PLS, and 0.75 (PLS 0.67) over 10 random 75/25 splits.

Model fixes

• Reconstruction no longer swamps prediction. The reconstruction error was summed over all predictors while the regression error was a mean, so with p spectral bands the model optimised reconstruction about p times harder than prediction (176 vs 0.68 for p = 200). The reconstruction error is now averaged over predictors.
• The target is standardised inside the model (mean and SD stored with it), so the balance of the loss terms no longer depends on the units of y. Predictions are returned in the original units.
• The regression head reads the latent mean, in training and at prediction. It was trained on a random latent sample while predicting from the mean.
• Early stopping watches the prediction loss (val_y_loss). It watched the total loss, dominated by reconstruction and computed on a random latent sample. With validation data, the learning rate is also halved when the prediction loss stalls (reduce_lr = TRUE).
• Stable defaults: beta_kl = 1e-3 (was 1), lr = 5e-4 (was 1e-3) and an L2 weight penalty l2 = 1e-4 (new).
• vae_predict() and vae_encode() are explicitly deterministic and use the latent mean. They previously relied on reticulate’s 0-based indexing of a Python tuple to pick it.
• A single hidden layer (hidden_enc = 64L) raised 'int' object is not iterable; it now works.

New

• vae_fit() scales the predictors and stores the scaling in the model, so the same spectra can be passed to vae_fit() and vae_predict(). The default scale_x = "global" centres each band and divides all bands by one overall SD, which keeps the shape of the spectrum; "band" scales each band by its own SD (this inflated noisy bands and occasionally gave very poor fits on spectra), "none" leaves X as given.
• Early stopping and learning-rate reduction wait for a warm-up of patience epochs, so a lucky early epoch can no longer end training before the model has learned (this happened in about one fit in ten).
• validation_split holds out training rows for early stopping when no X_val is given.
• The predictor and target scaling, the column names and a “fitted” flag are stored in the model (as non-trainable weights), so they are saved with it. vae_predict() and vae_encode() refuse an unfitted model, match columns by name when both sides have names, and accept a numeric vector for a single sample.
• vae_install() installs TensorFlow and Keras in the "r-soilvae" virtualenv (or a conda environment), which soilVAE then uses automatically in new sessions. vae_available() checks the backend without ever installing or downloading anything. Examples now run whenever the backend is available instead of never (\dontrun).
• vae_save() / vae_load(): save and reload a fitted model (architecture, weights, scaling and column names) in one .keras file.
• vae_history(): per-epoch training and validation losses.
• vae_build(seed = ) makes a fit reproducible; tune_vae_train_val() seeds each configuration (seed + i), frees memory between them and records l2.
• Every input is checked before the model is changed: finite numeric X and y (also X_val/y_val, given together), matching lengths, ranges of the vae_build() and vae_fit() arguments. select_best_from_grid() skips configurations with missing metrics.
• Requires TensorFlow >= 2.16 (Keras 3) and reticulate >= 1.41.
• A test suite (tests/testthat, 24 tests) covering each fix and feature.

Data

• The example dataset datsoilspc (redistributed from the ‘soilspec’ package) is removed. The vignette and README now use NIRsoil, loaded from the prospectr package at run time, so soilVAE distributes no data.

soilVAE 0.1.10

• Documentation-only release: declare the copyright holder of the bundled example dataset datsoilspc (redistributed unmodified from the ‘soilspec’ package by Alexandre M. J.-C. Wadoux, MIT license) in DESCRIPTION (cph role), LICENSE, inst/COPYRIGHTS, and in the dataset documentation (\source, \references).
• Added dataset provenance notes to the README and the vignette.
• No changes to code or functionality.

soilVAE 0.1.9

• Fixed vignette build to include prebuilt vignette index forced
• Minor documentation improvements