void dotrain(mdl_t *mdl) {
// Check if the user requested the trainer list. If this is not the
// case, search the trainer.
if (!strcmp(mdl->opt->algo, "list")) {
info("Available training algorithms:\n");
for (int i = 0; i < trn_cnt; i++)
info("\t%s\n", trn_lst[i].name);
exit(EXIT_SUCCESS);
}
int trn;
for (trn = 0; trn < trn_cnt; trn++)
if (!strcmp(mdl->opt->algo, trn_lst[trn].name))
break;
if (trn == trn_cnt)
fatal("unknown algorithm '%s'", mdl->opt->algo);
// Load a previous model to train again if specified by the user.
if (mdl->opt->model != NULL) {
info("* Load previous model\n");
FILE *file = fopen(mdl->opt->model, "r");
if (file == NULL)
pfatal("cannot open input model file");
mdl_load(mdl, file);
}
// Load the pattern file. This will unlock the database if previously
// locked by loading a model.
if (mdl->opt->pattern != NULL) {
info("* Load patterns\n");
FILE *file = fopen(mdl->opt->pattern, "r");
if (file == NULL)
pfatal("cannot open pattern file");
rdr_loadpat(mdl->reader, file);
fclose(file);
qrk_lock(mdl->reader->obs, false);
}
// Load the training data. When this is done we lock the quarks as we
// don't want to put in the model, informations present only in the
// devlopment set.
info("* Load training data\n");
FILE *file = stdin;
if (mdl->opt->input != NULL) {
file = fopen(mdl->opt->input, "r");
if (file == NULL)
pfatal("cannot open input data file");
}
mdl->train = rdr_readdat(mdl->reader, file, true);
if (mdl->opt->input != NULL)
fclose(file);
qrk_lock(mdl->reader->lbl, true);
qrk_lock(mdl->reader->obs, true);
if (mdl->train == NULL || mdl->train->nseq == 0)
fatal("no train data loaded");
// If present, load the development set in the model. If not specified,
// the training dataset will be used instead.
if (mdl->opt->devel != NULL) {
info("* Load development data\n");
FILE *file = fopen(mdl->opt->devel, "r");
if (file == NULL)
pfatal("cannot open development file");
mdl->devel = rdr_readdat(mdl->reader, file, true);
fclose(file);
}
// Initialize the model. If a previous model was loaded, this will be
// just a resync, else the model structure will be created.
if (mdl->theta == NULL)
info("* Initialize the model\n");
else
info("* Resync the model\n");
mdl_sync(mdl);
// Display some statistics as we all love this.
info("* Summary\n");
info(" nb train: %d\n", mdl->train->nseq);
if (mdl->devel != NULL)
info(" nb devel: %d\n", mdl->devel->nseq);
info(" nb labels: %zu\n", mdl->nlbl);
info(" nb blocks: %zu\n", mdl->nobs);
info(" nb features: %zu\n", mdl->nftr);
// And train the model...
info("* Train the model with %s\n", mdl->opt->algo);
uit_setup(mdl);
trn_lst[trn].train(mdl);
uit_cleanup(mdl);
// If requested compact the model.
if (mdl->opt->compact) {
const size_t O = mdl->nobs;
const size_t F = mdl->nftr;
info("* Compacting the model\n");
mdl_compact(mdl);
info(" %8zu observations removed\n", O - mdl->nobs);
info(" %8zu features removed\n", F - mdl->nftr);
}
// And save the trained model
info("* Save the model\n");
file = stdout;
if (mdl->opt->output != NULL) {
file = fopen(mdl->opt->output, "w");
if (file == NULL)
pfatal("cannot open output model");
}
mdl_save(mdl, file);
if (mdl->opt->output != NULL)
fclose(file);
info("* Done\n");
}
/*******************************************************************************
* Updating
******************************************************************************/
static void doupdt(mdl_t *mdl) {
// Load input model file
info("* Load model\n");
if (mdl->opt->model == NULL)
fatal("no model file provided");
FILE *Min = fopen(mdl->opt->model, "r");
if (Min == NULL)
pfatal("cannot open model file %s", mdl->opt->model);
mdl_load(mdl, Min);
fclose(Min);
// Open patch file
info("* Update model\n");
FILE *fin = stdin;
if (mdl->opt->input != NULL) {
fin = fopen(mdl->opt->input, "r");
if (fin == NULL)
pfatal("cannot open update file");
}
int nline = 0;
while (!feof(fin)) {
char *raw = rdr_readline(fin);
if (raw == NULL)
break;
char *line = raw;
nline++;
// First we split the line in space separated tokens. We expect
// four of them and skip empty lines.
char *toks[4];
int ntoks = 0;
while (ntoks < 4) {
while (isspace(*line))
line++;
if (*line == '\0')
break;
toks[ntoks++] = line;
while (*line != '\0' && !isspace(*line))
line++;
if (*line == '\0')
break;
*line++ = '\0';
}
if (ntoks == 0) {
free(raw);
continue;
} else if (ntoks != 4) {
fatal("invalid line at %d", nline);
}
// Parse the tokens, the first three should be string maping to
// observations and labels and the last should be the weight.
uint64_t obs = none, yp = none, y = none;
obs = qrk_str2id(mdl->reader->obs, toks[0]);
if (obs == none)
fatal("bad on observation on line %d", nline);
if (strcmp(toks[1], "#")) {
yp = qrk_str2id(mdl->reader->lbl, toks[1]);
if (yp == none)
fatal("bad label <%s> line %d", toks[1], nline);
}
y = qrk_str2id(mdl->reader->lbl, toks[2]);
if (y == none)
fatal("bad label <%s> line %d", toks[2], nline);
double wgh = 0.0;
if (sscanf(toks[3], "%lf", &wgh) != 1)
fatal("bad weight on line %d", nline);
const uint32_t Y = mdl->nlbl;
if (yp == none) {
double *w = mdl->theta + mdl->uoff[obs];
w[y] = wgh;
} else {
double *w = mdl->theta + mdl->boff[obs];
w[yp * Y + y] = wgh;
}
free(raw);
}
if (mdl->opt->input != NULL)
fclose(fin);
// If requested compact the model.
if (mdl->opt->compact) {
const uint64_t O = mdl->nobs;
const uint64_t F = mdl->nftr;
info("* Compacting the model\n");
mdl_compact(mdl);
info(" %8"PRIu64" observations removed\n", O - mdl->nobs);
info(" %8"PRIu64" features removed\n", F - mdl->nftr);
}
// And save the updated model
info("* Save the model\n");
FILE *file = stdout;
if (mdl->opt->output != NULL) {
file = fopen(mdl->opt->output, "w");
if (file == NULL)
pfatal("cannot open output model");
}
mdl_save(mdl, file);
if (mdl->opt->output != NULL)
fclose(file);
info("* Done\n");
}
