All fst4 CLI settings in one place: one field per command-line flag (plus a few derived/internal fields such as data_resolved), populated by CLI11 in add_options() and consumed by the factory functions and run_pipeline() below. Unless noted, string fields default to "" (flag omitted / feature off); concrete defaults live in the Config() constructor below. More...

Public Member Functions
bool	detailed_report () const
	true for --report detailed (the research sections; standard output is unchanged by them)
bool	data_is_stdin () const
	true for --data - (dataset piped in on standard input)
bool	data_free () const
const std::string &	data_path () const
	what the readers open: the stdin buffer when --data -, the given path otherwise
bool	has_missing_value () const
	true when --missing-value was given (NaN-off sentinel, self-compare)
bool	has_csv_missing_to () const
	true when --csv-missing-to was given (NaN-off sentinel, self-compare)
bool	is_filter_criterion () const
	true for the five built-in statistical-model filters (no classifier training)
bool	is_wrapper_criterion () const
	true for any "wrapper-*" criterion (classifier- or regressor-based)
bool	is_regression_criterion () const
	true for the two regression wrappers (dependent variable = the class label, "regression simulating classification")
bool	has_wrapper_metric () const
	false = the plain-accuracy path; true when --wrapper-metric/--class-weights select the metric palette
bool	nested_splits () const
	true when --splitter2 defines an inner (depth-1) splitting stack
bool	is_bb_search () const
bool	is_sequential_search () const
	true for the sequential-family searches (sfs/sbs/sffs/sbfs/sfrs/sbrs/os/dos) – the ones driven through an EVALUATOR/--evaluator

Public Attributes
std::string	data
	--data: path to the data file (TRN/ARFF/CSV), or "-" to read the dataset from standard input (then --format is required); "" when --features is used instead (data-free entry)
std::string	data_resolved
	--data -: the temporary file backing the buffered stdin; "" = --data is itself a path. The writers keep echoing --data, so the reported invocation stays "-" (and byte-stable)
std::string	format
	--format: data file format (trn\|arff\|csv); "" = infer from extension
std::string	csv_class
	CSV class column: "" = last column; integer = 0-based file column index; otherwise = header column name (implies a header row).
std::string	csv_delimiter
	auto \| comma \| semicolon \| tab (auto = detected from the first line; decimal commas accepted when delimiter != comma)
std::string	csv_header
	auto \| yes \| no (auto = any non-numeric non-class field in line 1 means header)
double	csv_missing_to
	NaN = off; else: the CSV loader maps missing FEATURE tokens to this code and keeps the row (pair with --missing-value <same code> for imputation; void scaler only; missing class labels still skip).
std::string	splitter
	outer splitter (splitting depth 0)
std::string	splitter2
	inner splitter (splitting depth 1); "" = single-level
std::string	scaler
	--scaler: data normalization – void (default, none) \| to01 (per-feature to [0,1]) \| white (whitening: zero mean, unit variance)
std::string	scaler_fit
	all (default: fit on the whole dataset before splitting) \| train (per-split: fit on each depth-0 TRAIN part only, so no test-part statistics leak into the transform)
std::string	criterion
	--criterion: feature-selection criterion name (a normal-/multinomial-model filter or a wrapper-* classifier/regressor criterion; see --list for the full registry with validity notes)
std::string	knn_distance
	--knn-distance: kNN distance metric – L1 \| L2 (default) \| Lp (p=3/2, compile-time menu)
std::string	search
	--search: search algorithm name (bif, sfs/sbs, sffs/sbfs, sfrs/sbrs, os, dos, exhaustive, montecarlo, bb-basic/bb-improved/bb-fast/bb-bbpp)
std::string	search2
	seeding pre-search ("bif"): its result subset seeds the main search (demo25t/33)
std::string	svm_kernel
	linear \| rbf \| poly \| sigmoid (default rbf; LINEAR/SIGMOID may freeze LibSVM on some data)
std::string	svm_optimize
	none \| pre \| alternating: LibSVM C/gamma/coef0 grid optimization (demo12t / demo23)
std::string	regr_multipliers
	comma list: kernel-multiplier sweep, best-by-train wins (demo11reg)
std::string	evaluator
	sequential step object: straight \| hybrid (filter prefilter, demo53) \| ensemble (voting, demo50) \| sampled (budgeted stochastic step)
std::string	hybrid_prefilter
	filter criterion pre-filtering candidates (with --evaluator hybrid)
std::string	ensemble_knn_ks
	comma list of kNN k values forming the voting ensemble (with --evaluator ensemble)
std::string	step_sampler
	sampled evaluator proposal policy: softmax \| uniform \| topk
std::string	step_stats
	sampled evaluator per-feature statistic: zcontrast \| rawmean
std::string	classify_samples
	file of whitespace-separated samples to classify on the selected subspace (demo63)
std::string	validate
	comma list of classifiers for final depth-0 validation; "" = none; "same" = the wrapper criterion's classifier
std::string	wrapper_metric
	wrapper criterion metric beyond accuracy: balanced-accuracy \| f1 \| f1-macro \| auc (classification) \| r2 \| neg-rmse \| neg-mae (regression); "" = accuracy (default)
std::string	class_weights
	comma list of per-class cost weights, class-index order (classification wrappers); "" = unweighted
unsigned int	f1_positive
	class index scored as positive by --wrapper-metric f1 (two-class data; default 1)
std::string	tracker
	dupless (default) \| regularizer \| stabileval \| feature-stats
std::string	tracker_attach
	"" = auto (search object; regularizer attaches to the evaluator per demo choreography)
std::string	regularize_by
	secondary criterion for the regularization sweep: subsetsize \| weights
std::string	regularize_margins
	comma list of margins for optimize_within_margin
std::string	feature_costs
	comma list of per-feature costs (with --regularize-by weights)
std::string	output
	--output: result format on stdout – text (default) \| json \| csv; with json/csv all search progress goes to stderr
std::string	verbosity
	--verbosity: search progress verbosity – silent \| normal (default) \| detailed
std::string	progress_mode
	search progress delivery: text (default) \| json (one line per improved solution) \| quiet (observer swallows progress)
std::string	compare_config
	config file of a second scenario: run it too and report cross-scenario tracker similarity (the demo55 protocol)
double	missing_value
	NaN = off; else: substitute this exact value by the per-feature average of the available values – dataset-wide by default, the depth-0 TRAIN part's under --scaler-fit train (demo62; void scaler only).
unsigned int	cv_folds
	--cv-folds: number of outer cross-validation folds (with --splitter cv; default 3)
unsigned int	holdout_train
	--holdout-train: percentage of data used for outer training (with --splitter holdout; default 70)
unsigned int	rr_splits
	--rr-splits: outer randrand/randfix run count (number of random splits)
unsigned int	rr_train
	--rr-train: % of samples drawn for outer training
unsigned int	rr_test
	--rr-test: % of samples drawn for outer testing (disjoint from training)
unsigned int	cv2_folds
	--splitter2-cv-folds: number of inner cross-validation folds (with --splitter2 cv; default 3)
unsigned int	holdout2_train
	--splitter2-holdout-train: inner training percentage (with --splitter2 holdout; default 70)
unsigned int	rr2_splits
	--splitter2-rr-splits: inner randrand/randfix run count – the --splitter2 twin of rr_splits
unsigned int	rr2_train
	--splitter2-rr-train: % of samples drawn for inner training – the --splitter2 twin of rr_train
unsigned int	rr2_test
	--splitter2-rr-test: % of samples drawn for inner testing – the --splitter2 twin of rr_test
unsigned int	knn_k
	--knn-k: k for the kNN classifier (with --criterion wrapper-knn; default 3)
double	ridge_lambda
	ridge regression L2 weight penalty (default 0.01; bias unpenalized)
std::string	extern_cmd
	external-tool bridge command (with --criterion wrapper-extern or extern-value); USER-SUPPLIED CODE EXECUTION — see classifier_extern.hpp / criterion_extern_value.hpp
bool	extern_monotonic
	the user's assertion that the extern-value measure is cardinality-monotonic (admits bb-* searches; optimality rests on the assertion)
bool	extern_no_data
	suppress the extern-value DATA block: the tool is a pure set function over index sets
unsigned int	features_n
	--features N: data-free entry — an abstract universe of N features, no --data at all (extern-value --extern-no-data only; "Tier 2")
double	linsvm_c
	linear SVM regularization C (default 1.0)
std::string	linsvm_loss
	linear SVM loss: l2 (squared hinge, default) \| l1 (hinge)
unsigned int	linsvm_maxiter
	linear SVM dual CD sweep cap (default 1000)
double	linsvm_tol
	linear SVM projected-gradient stopping threshold (default 0.1)
std::string	rbfsvm_optimize
	none \| pre \| alternating: native RBF SVM C/gamma optimization (mirrors --svm-optimize)
double	rbfsvm_c
	RBF SVM regularization C (default 1.0).
double	rbfsvm_gamma
	RBF width on standardized features (default 0 = auto: 1/d).
std::string	rbfsvm_loss
	RBF SVM loss: l1 (hinge, the C-SVC correspondence, default) \| l2 (squared hinge).
unsigned int	rbfsvm_maxiter
	RBF SVM kernel dual CD sweep cap (default 1000).
double	rbfsvm_tol
	RBF SVM projected-gradient stopping threshold (default 0.1).
unsigned int	rf_trees
	Random Forest wrapper: number of trees (default 100).
unsigned int	rf_mtry
	features tried per node; 0 = floor(sqrt(subset size))
unsigned int	rf_max_depth
	Random Forest tree depth cap; 0 = unlimited.
unsigned int	rf_min_leaf
	minimum samples per leaf (default 1)
unsigned int	search2_d
	seed subset size for --search2 (0 = use --target-d)
double	svm_c
	--svm-c: LibSVM regularization C (with --criterion wrapper-svm); 0 = LibSVM library default, or the value found by --svm-optimize
double	svm_gamma
	--svm-gamma: LibSVM RBF/poly/sigmoid kernel width (with --criterion wrapper-svm); 0 = LibSVM library default, or the value found by --svm-optimize
double	svm_coef0
	--svm-coef0: LibSVM poly/sigmoid kernel bias term (with --criterion wrapper-svm); 0 = LibSVM library default, or the value found by --svm-optimize
double	logreg_lambda
	logistic regression L2 weight penalty (biases unpenalized)
unsigned int	logreg_maxiter
	logistic regression Newton iteration cap
double	logreg_tol
	logistic regression gradient-norm stopping tolerance
double	regr_multiplier
	weighted-neighbours kernel multiplier (0 = library default)
unsigned int	hybrid_keep_perc
	hybrid evaluator: percentage of prefiltered candidates kept (demo53: 50)
unsigned int	hybrid_keep_max
	hybrid evaluator: hard cap on kept candidates
unsigned int	step_cap
	sampled evaluator: max candidate evaluations per forward step (0 = no cap)
unsigned int	step_cap_backward
	sampled evaluator: max candidate evaluations per backward step (0 = full sweeps, exact floating correction)
double	step_cap_frac
	sampled evaluator: constant-percentage schedule, cap = ceil(frac*pool) (0 = off, use step_cap); forward only
bool	step_frozen
	sampled evaluator: freeze statistics after warm-up (frozen-ranking ablation)
double	step_explore
	sampled evaluator: exploration-floor fraction of each proposal batch
double	step_tau
	sampled evaluator: softmax temperature (0 = robust automatic scale)
unsigned int	step_decay
	sampled evaluator: per-feature statistics averaging horizon (evaluations)
unsigned long	warmup_probes
	sampled evaluator: random probes seeding the statistics before the search (0 = off)
unsigned int	warmup_card
	sampled evaluator: warm-up probe cardinality
unsigned int	restarts_nonimproving
	random-restart protocol: stop after N consecutive non-improving runs (0 = off; demo22)
unsigned long	restarts_max
	random-restart safety cap on total runs (0 = unlimited)
bool	bias_estimate
	estimate classifier bias on the selected subset (demo56)
unsigned int	target_d
	target subset size (--target-d); 0 = optimize subset size (wrapper criteria only – monotonic filter criteria and B&B/os require a nonzero value)
unsigned int	os_delta
	oscillation cycle depth (--search os; default 1)
unsigned int	dos_delta
	oscillation cycle depth (--search dos; default 3)
unsigned int	sffs_delta
	floating-frontier bound: SFFS/SBFS stop at target-d +/- delta (0 = classic full frontier)
unsigned int	generalization
	g-tuples per sequential step (demo21); 1 = classic
unsigned long long	mc_trials
	montecarlo stopping: max trials (0 = unlimited if mc_seconds set)
unsigned long	mc_seconds
	montecarlo stopping: wall-clock budget (0 = none)
unsigned int	mc_card_min
unsigned int	mc_card_max
	montecarlo probe cardinality range upper bound (with mc_card_min, 0,0 = full range)
double	mc_card_prob
	montecarlo per-feature inclusion probability (0 = off)
double	bb_gamma
	FBB prediction optimism (default 1.0).
unsigned int	bb_delta
	FBB minimum prediction evidence (default 1).
unsigned long	tracker_capacity
	0 = type default (dupless: top_k; others: unlimited)
double	tracker_margin
	dupless-family inclusion margin (0 = only improving results)
int	daf_rank
	DAF ranking variant to extract from feature-stats (-1 = off; 0\|1).
unsigned int	daf_count
	number of top-DAF features to report (0 = all)
unsigned int	daf_validate_every
	validate incrementally built top-DAF subsets every K steps (0 = off)
unsigned int	daf_pairs
	pairwise DAF statistics over the top-K DAF0 features (0 = off)
unsigned int	daf_pairs_top
	number of top pairs (by pair-DAF0) to report (0 = all)
std::string	report
	standard (default) \| detailed (adds the research sections: curve_over_d, ROC/PR, confusion, tuned parameters)
std::string	dump_scores
	file to receive per-holdout-sample true/predicted labels + per-class scores of the final subset (the first --validate classifier); "" = off
std::string	curves_svg
	file to receive a dependency-free SVG rendering of the --report detailed curves (ROC, PR, criterion-vs-holdout over d); "" = off
unsigned int	threads
	number of worker threads for candidate evaluation; 1 = sequential evaluation, deterministic ordering (default 1)
unsigned int	top_k
	number of best tracked subsets to report (default 5); also the dupless tracker's default capacity
unsigned int	seed
	random seed (--seed); affects randomized components only, e.g. the initial subset of --search os (default 1)
bool	stable_output
	omit non-deterministic output (timings) so results are byte-comparable across runs
bool	repeat_over_splits
	run the search once per outer split (trial protocol, demo54)

Detailed Description

All fst4 CLI settings in one place: one field per command-line flag (plus a few derived/internal fields such as data_resolved), populated by CLI11 in add_options() and consumed by the factory functions and run_pipeline() below. Unless noted, string fields default to "" (flag omitted / feature off); concrete defaults live in the Config() constructor below.

Member Function Documentation

◆ data_free()

bool fst4_cli::Config::data_free ( ) const

inline

true for --features N: no data file exists, the universe is abstract

◆ is_bb_search()

bool fst4_cli::Config::is_bb_search ( ) const

inline

true for any of the four Branch & Bound variants (--search bb-*)

Member Data Documentation

◆ mc_card_min

unsigned int fst4_cli::Config::mc_card_min

< montecarlo probe cardinality range lower bound (with mc_card_max, 0,0 = full range)

The documentation for this struct was generated from the following file:

/Users/psms/dev/devtest1/fst4.cpp