X-Git-Url: https://pintos-os.org/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=src%2Flanguage%2Fstats%2Fglm.q;h=d3a286090fe0ad6bf625ba546b5500fda38a6a16;hb=9f255a7d7f35f2bbde3e0f79886ea17931933d35;hp=39e4f1c5d207ab57c0d9735361bfc66cd36242fa;hpb=fa1cc7513bbb306fb00503007575db8ffb76602e;p=pspp-builds.git diff --git a/src/language/stats/glm.q b/src/language/stats/glm.q index 39e4f1c5..d3a28609 100644 --- a/src/language/stats/glm.q +++ b/src/language/stats/glm.q @@ -1,5 +1,5 @@ /* PSPP - a program for statistical analysis. - Copyright (C) 2007 Free Software Foundation, Inc. + Copyright (C) 2007, 2009 Free Software Foundation, Inc. This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -37,6 +37,7 @@ #include #include #include +#include #include #include #include @@ -47,13 +48,12 @@ #include "xalloc.h" #include "gettext.h" -#define GLM_LARGE_DATA 10000 - /* (headers) */ /* (specification) "GLM" (glm_): *dependent=custom; + design=custom; by=varlist; with=varlist. */ @@ -61,6 +61,7 @@ /* (functions) */ static struct cmd_glm cmd; + /* Moments for each of the variables used. */ @@ -84,14 +85,12 @@ static const struct variable **v_dependent; */ static size_t n_dependent; -#if 0 -/* - Return value for the procedure. - */ -static int pspp_glm_rc = CMD_SUCCESS; -#else +size_t n_inter; /* Number of interactions. */ +size_t n_members; /* Number of memebr variables in an interaction. */ + +struct interaction_variable **interactions; + int cmd_glm (struct lexer *lexer, struct dataset *ds); -#endif static bool run_glm (struct casereader *, struct cmd_glm *, @@ -120,7 +119,60 @@ cmd_glm (struct lexer *lexer, struct dataset *ds) free (v_dependent); return ok ? CMD_SUCCESS : CMD_FAILURE; } +static int +parse_interactions (struct lexer *lexer, const struct variable **interaction_vars, int n_members, + int max_members, struct dataset *ds) +{ + if (lex_match (lexer, '*')) + { + if (n_members > max_members) + { + max_members *= 2; + xnrealloc (interaction_vars, max_members, sizeof (*interaction_vars)); + } + interaction_vars[n_members] = parse_variable (lexer, dataset_dict (ds)); + parse_interactions (lexer, interaction_vars, n_members++, max_members, ds); + } + return n_members; +} +/* Parser for the design subcommand. */ +static int +glm_custom_design (struct lexer *lexer, struct dataset *ds, + struct cmd_glm *cmd UNUSED, void *aux UNUSED) +{ + size_t n_inter = 0; + size_t n_allocated = 2; + size_t n_members; + struct variable **interaction_vars; + struct variable *this_var; + interactions = xnmalloc (n_allocated, sizeof (*interactions)); + + while (lex_token (lexer) != T_STOP && lex_token (lexer) != '.') + { + this_var = parse_variable (lexer, dataset_dict (ds)); + if (lex_match (lexer, '(')) + { + lex_force_match (lexer, ')'); + } + else if (lex_match (lexer, '*')) + { + n_members = 1; + interaction_vars = xnmalloc (2 * n_inter, sizeof (*interaction_vars)); + n_members = parse_interactions (lexer, interaction_vars, 1, 2 * n_inter, ds); + if (n_allocated < n_inter) + { + n_allocated *= 2; + xnrealloc (interactions, n_allocated, sizeof (*interactions)); + } + interactions [n_inter - 1] = + interaction_variable_create (interaction_vars, n_members); + n_inter++; + free (interaction_vars); + } + } + return 1; +} /* Parser for the dependent sub command */ static int glm_custom_dependent (struct lexer *lexer, struct dataset *ds, @@ -152,65 +204,26 @@ glm_custom_dependent (struct lexer *lexer, struct dataset *ds, model. That means the dependent variable is in there, too. */ static void -coeff_init (pspp_linreg_cache * c, struct design_matrix *cov) +coeff_init (pspp_linreg_cache * c, const struct design_matrix *cov) { c->coeff = xnmalloc (cov->m->size2, sizeof (*c->coeff)); c->n_coeffs = cov->m->size2 - 1; pspp_coeff_init (c->coeff, cov); } -/* Encode categorical variables. - Returns number of valid cases. */ -static int -data_pass_one (struct casereader *input, - const struct variable **vars, size_t n_vars, - struct moments_var **mom) -{ - int n_data; - struct ccase c; - size_t i; - - for (i = 0; i < n_vars; i++) - { - mom[i] = xmalloc (sizeof (*mom[i])); - mom[i]->v = vars[i]; - mom[i]->mean = xmalloc (sizeof (*mom[i]->mean)); - mom[i]->variance = xmalloc (sizeof (*mom[i]->mean)); - mom[i]->weight = xmalloc (sizeof (*mom[i]->weight)); - mom[i]->m = moments1_create (MOMENT_VARIANCE); - if (var_is_alpha (vars[i])) - cat_stored_values_create (vars[i]); - } - n_data = 0; - for (; casereader_read (input, &c); case_destroy (&c)) - { - /* - The second condition ensures the program will run even if - there is only one variable to act as both explanatory and - response. - */ - for (i = 0; i < n_vars; i++) - { - const union value *val = case_data (&c, vars[i]); - if (var_is_alpha (vars[i])) - cat_value_update (vars[i], val); - else - moments1_add (mom[i]->m, val->f, 1.0); - } - n_data++; - } - casereader_destroy (input); - for (i = 0; i < n_vars; i++) - { - if (var_is_numeric (mom[i]->v)) - { - moments1_calculate (mom[i]->m, mom[i]->weight, mom[i]->mean, - mom[i]->variance, NULL, NULL); - } - } - - return n_data; +static pspp_linreg_cache * +fit_model (const struct covariance_matrix *cov, + const struct variable *dep_var, + const struct variable ** indep_vars, + size_t n_data, size_t n_indep) +{ + pspp_linreg_cache *result = NULL; + result = pspp_linreg_cache_alloc (dep_var, indep_vars, n_data, n_indep); + coeff_init (result, covariance_to_design (cov)); + pspp_linreg_with_cov (cov, result); + + return result; } static bool @@ -218,29 +231,27 @@ run_glm (struct casereader *input, struct cmd_glm *cmd, const struct dataset *ds) { - pspp_linreg_cache *model = NULL; - size_t i; - size_t j; - int n_indep = 0; - struct ccase c; + casenumber row; const struct variable **indep_vars; const struct variable **all_vars; - struct design_matrix *X; - struct moments_var **mom; - struct casereader *reader; - casenumber row; + int n_indep = 0; + pspp_linreg_cache *model = NULL; + pspp_linreg_opts lopts; + struct ccase *c; + size_t i; size_t n_all_vars; size_t n_data; /* Number of valid cases. */ + struct casereader *reader; + struct covariance_matrix *cov; - pspp_linreg_opts lopts; - - if (!casereader_peek (input, 0, &c)) + c = casereader_peek (input, 0); + if (c == NULL) { casereader_destroy (input); return true; } - output_split_file_values (ds, &c); - case_destroy (&c); + output_split_file_values (ds, c); + case_unref (c); if (!v_dependent) { @@ -265,70 +276,44 @@ run_glm (struct casereader *input, all_vars[i + n_dependent] = cmd->v_by[i]; } n_indep = cmd->n_by; - mom = xnmalloc (n_all_vars, sizeof (*mom)); - reader = casereader_clone (input); reader = casereader_create_filter_missing (reader, indep_vars, n_indep, MV_ANY, NULL, NULL); reader = casereader_create_filter_missing (reader, v_dependent, 1, MV_ANY, NULL, NULL); - n_data = data_pass_one (casereader_clone (reader), - (const struct variable **) all_vars, n_all_vars, - mom); - if ((n_data > 0) && (n_indep > 0)) + if (n_indep > 0) { for (i = 0; i < n_all_vars; i++) if (var_is_alpha (all_vars[i])) cat_stored_values_create (all_vars[i]); - X = - covariance_matrix_create (n_all_vars, - (const struct variable **) all_vars); + cov = covariance_matrix_init (n_all_vars, all_vars, ONE_PASS, PAIRWISE, MV_ANY); + reader = casereader_create_counter (reader, &row, -1); - for (; casereader_read (reader, &c); case_destroy (&c)) + + for (i = 0; i < n_inter; i++) + if (var_is_alpha (interaction_get_variable (interactions[i]))) + cat_stored_values_create (interaction_get_variable (interactions[i])); + covariance_interaction_set (cov, interactions, 1); + for (; (c = casereader_read (reader)) != NULL; case_unref (c)) { /* Accumulate the covariance matrix. - */ - for (i = 0; i < n_all_vars; ++i) - { - const struct variable *v = all_vars[i]; - const union value *val_v = case_data (&c, v); - if (var_is_alpha (all_vars[i])) - cat_value_update (all_vars[i], val_v); - for (j = i; j < n_all_vars; j++) - { - const struct variable *w = all_vars[j]; - const union value *val_w = case_data (&c, w); - covariance_pass_two (X, *mom[i]->mean, *mom[j]->mean, - (double) n_data, - v, w, val_v, val_w); - } - } + */ + covariance_matrix_accumulate (cov, c, interactions, 1); + n_data++; } - model = pspp_linreg_cache_alloc (v_dependent[0], indep_vars, n_data, n_indep); - /* - For large data sets, use QR decomposition. - */ - if (n_data > sqrt (n_indep) && n_data > GLM_LARGE_DATA) + covariance_matrix_compute (cov); + for (i = 0; i < n_dependent; i++) { - model->method = PSPP_LINREG_QR; + model = fit_model (cov, v_dependent[i], indep_vars, n_data, n_indep); + pspp_linreg_cache_free (model); } - coeff_init (model, X); - pspp_linreg_with_cov (X, model); + casereader_destroy (reader); - for (i = 0; i < n_all_vars; i++) - { - moments1_destroy (mom[i]->m); - free (mom[i]->mean); - free (mom[i]->variance); - free (mom[i]->weight); - free (mom[i]); - } - free (mom); - covariance_matrix_destroy (X); + covariance_matrix_destroy (cov); } else { @@ -336,7 +321,6 @@ run_glm (struct casereader *input, } free (indep_vars); free (lopts.get_indep_mean_std); - pspp_linreg_cache_free (model); casereader_destroy (input); return true;