/* File name: bagging.c Created by: Ljubomir Buturovic Created: 07/31/2002 Purpose: implement bagging (committee) variants of common pattern clssification algorithms. */ /* Copyright 2004 Ljubomir J. Buturovic Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. */ static char rcsid[] = "$Id: bagging.c,v 1.71 2006/04/04 18:18:34 ljubomir Exp $"; #include #include #include #include #include #include "dataset.h" #include "bagging.h" #include "mlp.h" #include "svm.h" #include "pcl_svm.h" #include "lau.h" #include "lin.h" #include "lind.h" #include "parametric.h" #include "adaboost.h" /* Train a combined classifier using 'bagging' algorithm described in: Amanda J. C. Sharkey (Ed.), Combining Artificial Neural Nets, Chapter 2, Sections 2.2.2 and 3.2. Springer, London, 1999. 'nmodels' is the number of classifiers used for bagging; 'bag_size' is the size of each sampling dataset used to build an individual classifier. The sampling datasets are obtained by resampling the input 'dset'. Return array of 'nmodels' classifier structs (one per sampling dataset), representing the bagging classifier. The function stores the classifier in file 'fname'. In case of error, return NULL and set 'errc'. The 'problems' argument is specific to SVM bagging. It returns the set of SVM svm_problems structures used to build the bagging classifier, for subsequent free()-ing. The reason that the structures are not free()-ed here is that the returned models reference them, so the free() can only be done by the caller once the models are no longer needed. This function has been reimplemented as pcl_svm_learn() for SVM classifiers, and put into libpcl.a library, for external use. Consider reimplementing bagging for all classifiers in such manner. */ void **bagging(struct dataset *dset, FILE *outdev, int method, char *fname, unsigned int seed, int nmodels, int bag_size, void *options, void ***problems, int *errc, FILE *fdbg) { int i; int d; int nvec; int status; int mode; int *bnd; char *tname; /* temporary MLP file */ float **bag = (float **) 0; float **target; void *model; /* individual model */ void **models = (void **) 0; /* array of individual models, comprising the bagging classifier */ struct dataset *bag_set; struct mlp_options *mlp_optional; struct svm_parameter *parameters; struct svm_problem *problem; struct svm_problem **probs; FILE *fptr; model = (void *) 0; mlp_optional = (struct mlp_options *) 0; parameters = (struct svm_parameter *) 0; probs = (struct svm_problem **) 0; fptr = (FILE *) 0; d = dset->d; if (method == PALG_BAG_MLP) mlp_optional = (struct mlp_options *) options; else if (method == PALG_BAG_SVM) { parameters = (struct svm_parameter *) options; probs = malloc(nmodels*sizeof(struct svm_problem *)); } /* For each bagging subset, train a classifier and append the model to the output file. */ bag = malloc(bag_size*sizeof(float *)); bnd = calloc(dset->c, sizeof(int)); status = 0; tname = tempfile(); if (tname == (char *) 0) status = -1; if (status == 0) models = calloc(nmodels+1, sizeof((void **) 0)); if (fname && *fname) { fptr = fopen(fname, "w"); if (!fptr) status = -1; } for (i = 0; (i < nmodels) && (status == 0); i++) { nvec = resample(i, dset, bag_size, bag, bnd, fdbg); if (method == PALG_BAG_MLP) { target = mlp_target(dset->c, bnd); /* Train, then append the model to 'fname'. */ model = mlp_learn(mlp_optional->opt_method, bag, nvec, bnd, d, target, mlp_optional->nlayers, mlp_optional->npl, mlp_optional->itmax, mlp_optional->range, mlp_optional->eta, mlp_optional->mu, outdev, 0, tname, seed, errc, fdbg); mx_free((void **) target, nvec); if (model) { free(((struct mlp *) model)->fname); ((struct mlp *) model)->fname = strdup(fname); status = mlp_save(model, MLP_MODE_APPEND, i+1, 1.0); } else status = -1; } else if (method == PALG_BAG_SVM) { bag_set = dataset_lt(d, dset->c, bnd, nvec, (char **) 0, bag); problem = create_problem(bag_set); free(bag_set); probs[i] = problem; model = svm_train(problem, parameters); if (model) { status = save_svm(fptr, model, i+1, 1.0); if (status == 0) status = fflush(fptr); else { *errc = errno; fptr_close(fptr); } } else { status = -1; *errc = errno; } } else if ((method == PALG_BAG_LIN) || (method == PALG_BAG_PLC)) { bag_set = dataset_lt(d, dset->c, bnd, nvec, (char **) 0, bag); if (method == PALG_BAG_LIN) model = lind_learn(bag_set, errc, fdbg); else { /* TBD: make mode a parameter to bagging(). */ mode = WEIGHTED_COV; model = lin_learn(mode, bag_set, errc); } if (model) lin_write(fptr, nmodels, model, dset->c, d+1, i+1, 1.0); else status = -1; free(bag_set); } else if (method == PALG_BAG_PQC) { bag_set = dataset_lt(d, dset->c, bnd, nvec, (char **) 0, bag); model = pqc_learn(bag_set, errc); if (model) status = pqc_write(fptr, nmodels, model, d, dset->c, i+1, 1.0); else status = -1; mx_free((void **) bag_set->sigma, dset->c); free(bag_set->det); free(bag_set); } models[i] = model; } if (status == -1) { free(models); models = (void **) 0; } unlink(tname); free(tname); free(bag); free(bnd); fptr_close(fptr); if (problems) *problems = (void **) probs; return models; }