fit_abund_qrf now works when a partition is used; it was calling the internal result wrapper with an outdated argument order. Without a partition it now also returns predictors and metadata, as the other fit_abund_* functions do.
fit_abund_glm and fit_abund_gam no longer fail when hold_out_set is used (object 'train_set' not found).
fit_abund_net no longer fails when hold_out_set is used (nnet predictions require type = "raw").
fit_abund_dnn now evaluates hold_out_set correctly and no longer builds a mismatched default network when predictors_f is supplied (categorical predictors are ignored, with a warning).
fit_abund_cnn now works without a partition; using hold_out_set with CNN models now gives an informative error because it is not yet supported.
adm_predict now names the predictions of quantile regression forests qrf (they were named NA), and p_abund_pdp/adm_predict no longer rely on $feature_names, which recent xgboost versions do not provide.
adm_uncertainty now accepts ... (as documented), so the extra arguments needed for CNN models (x, y, rasters, sample_size, custom_architecture) can be passed.
Testing
New tests for fit_abund_* (all algorithms, including no-partition, hold-out and DNN/CNN paths), tune_abund_qrf, adm_transform (inverse and raster), adm_predict, adm_uncertainty, model_selection, data_abund_pdp, p_abund_pdp, p_abund_bpdp, generate_dnn_architecture, generate_cnn_architecture, res_calculate, family_selector, croppin_hood, cnn_make_samples and get_partition_samples.
The slow tune_abund_* tests (all but the XGB one and the error checks), the DNN/CNN prediction tests and the slowest data_abund_bpdp test are skipped on CRAN, and the vignette grids were reduced, to keep check time low.
The DNN/CNN tuning and prediction tests were shortened (smaller data, networks and epochs) and the heaviest ones are skipped on CRAN, reducing the test time from about 16 to 2 minutes.
The test-coverage workflow installs libtorch so the DNN/CNN tests run, and now uploads the covr report (cobertura.xml) to Codecov explicitly; previously the wrong file was uploaded and coverage was reported as 0.
Documentation and CRAN preparation
Documented previously undocumented arguments (hold_out_set, samples_list, weight_decay, optimizer, loss_function, framework, train_quantiles, eval_quantile, pred_quantile, sample_prop, na.rm, and the arguments of get_partition_samples) and removed the non-existent hold_out_evaluation argument from the XGB help pages.
grf and quantregForest were added to Suggests.
The DNN sections of the “Modeling workflow” vignette are only evaluated when torch is installed; vignette titles now match their index entries.
Declared global variables to avoid R CMD check notes; updated the package title, CITATION, and README links.
adm 0.0.2
fit_abund_qrf & tune_abund_qrf were implemented to construct Quantile Regression Random Forests models
p_abund_pdp was improved to depict exactly training range values when projection data are used #209
p_abund_pdp for GLM was fixed when polynomials are used #209
p_abund_bpdp for XBT was fixed when polynomials are used #209