void recbld(void)
{
int i;
unsigned char *r, *p;
if (donerec == 1)
return;
r = recdata;
for (i = 1; i <= *NF; i++) {
p = getsval(fldtab[i]);
while ((*r = *p++)) {
if (++r >= &recdata[recsize]) {
recsize += CHUNK;
growrec(&recdata, &recsize, recsize, &r, 1);
}
}
if (i < *NF)
for ((p = *OFS); (*r = *p++); ) {
if (++r >= &recdata[recsize]) {
recsize += CHUNK;
growrec(&recdata, &recsize,
recsize, &r, 1);
}
}
}
*r = '\0';
dprintf( ("in recbld FS=%o, recloc=%lo\n", **FS,
(long)recloc) );
recloc->tval = REC | STR | DONTFREE;
recloc->sval = record = recdata;
dprintf( ("in recbld FS=%o, recloc=%lo\n", **FS,
(long)recloc) );
dprintf( ("recbld = |%s|\n", record) );
donerec = 1;
}
int readrec(unsigned char **buf, int *bufsize, FILE *inf)
/* read one record into buf */
{
register int sep, c, k, m, n;
unsigned char *rr;
register int nrr;
wchar_t wc;
next(wc, *RS, n);
if ((sep = **RS) == 0) {
sep = '\n';
while ((c=getc(inf)) == '\n' && c != EOF) /* skip leading \n's */
;
if (c != EOF)
ungetc(c, inf);
}
if (*bufsize == 0)
growrec(buf, bufsize, CHUNK, NULL, 0);
for (rr = *buf, nrr = *bufsize; ; ) {
cont: for (; (c=getc(inf)) != sep && c != EOF; *rr++ = c)
if (--nrr < n + 3) {
growrec(buf, bufsize, *bufsize + CHUNK, &rr, 0);
nrr += CHUNK;
}
if (c != EOF) {
/*
* Note: This code does not restrict occurences of
* the multibyte sequence in RS to the start of an
* input character.
*/
for (m = 1; m < n; m++) {
if ((c = getc(inf)) == EOF || c != (*RS)[m]) {
for (k = 0; k < m; k++)
*rr++ = (*RS)[k];
nrr -= k;
if (c == EOF)
break;
*rr++ = c;
nrr--;
goto cont;
}
}
}
if (**RS == sep || c == EOF)
break;
if ((c = getc(inf)) == '\n' || c == EOF) /* 2 in a row */
break;
*rr++ = '\n';
*rr++ = c;
}
/*if (rr > *buf + *bufsize)
error(MM_ERROR, ":12:Input record `%.20s...' too long", *buf);*/
*rr = 0;
dprintf( ("readrec saw <%s>, returns %d\n", *buf, c == EOF
&& rr == *buf ? 0 : 1) );
return c == EOF && rr == *buf ? 0 : 1;
}
void
recbld(void)
{
int i;
register char *r, *p;
if (donefld == 0 || donerec == 1)
return;
r = record;
for (i = 1; i <= *NF; i++) {
p = getsval(fldtab[i]);
while (*r++ = *p++) {
if (r >= record+RECSIZE) {
size_t diff = r - record;
growrec();
r = &record[diff];
}
}
*(r-1) = **OFS;
}
*(r-1) = '\0';
dprintf("in recbld FS=%lo, recloc=%lo\n", (long)**FS, (long)recloc);
recloc->tval = STR | FLD;
dprintf("in recbld FS=%lo, recloc=%lo\n", (long)**FS, (long)recloc);
/*if (r > record+RECSIZE)
error(FATAL, "built giant record `%.20s...'", record);*/
dprintf("recbld = |%s|\n", record);
}
unsigned char *makerec(const unsigned char *data, int size)
{
if (!(recloc->tval & DONTFREE))
xfree(recloc->sval);
if (recsize < size)
growrec(&recdata, &recsize, size, NULL, 0);
record = recdata;
strcpy((char*)record, (char*)data);
recloc->sval = record;
recloc->tval = REC | STR | DONTFREE;
donerec = 1; donefld = 0;
return record;
}
void grow(tree_t* t, dataset_t* d){
int i;
/* Initialize root fields */
t->root = malloc(sizeof(node_t));
t->root->pos = FLT_EPSILON;
t->root->neg = FLT_EPSILON;
for(i=0; i<d->nex; i++){
if(t->valid[i]<=0)
continue;
if(d->target[i])
t->root->pos += d->weight[i];
else
t->root->neg += d->weight[i];
}
t->root->pos = min(1-FLT_EPSILON, t->root->pos);
t->root->neg = min(1-FLT_EPSILON, t->root->neg);
/* Recursively grow tree */
growrec(t, t->root, d, 0);
}
void growrec(tree_t* t, node_t* root, dataset_t* d, int depth){
split_t best;
int i,k,l,u;
node_t* first;
node_t* second;
evpair_t* b;
/* Stop if max depth is reached or node is pure */
if(depth>=t->maxdepth || root->pos <= FLT_EPSILON || root->neg <= FLT_EPSILON){
root->split=-1;
return;
}
/* Find the best split */
best = bestSplit(t,root,d);
/* Stop if no good split is left or the counts in one of the children are very small */
if (best.feature < 0 ||
(best.posleft <= FLT_EPSILON && best.negleft <= FLT_EPSILON) ||
(best.posright <= FLT_EPSILON && best.negright <= FLT_EPSILON)){
root->split=-1;
return;
}
/* Install the split */
root->split=best.feature;
root->threshold=best.threshold;
root->left=malloc(sizeof(node_t));
root->left->pos=best.posleft;
root->left->neg=best.negleft;
root->right=malloc(sizeof(node_t));
root->right->pos=best.posright;
root->right->neg=best.negright;
/* Mark the feature as used */
if(!d->cont[best.feature])
t->used[best.feature]=1;
b = d->feature[best.feature];
/* Find the first example whose value exceeds the threshold */
k = 0;
u = d->size[best.feature];
while (k < u) {
i = (k + u)/2;
if (b[i].value > best.threshold)
u = i;
else
k = i + 1;
}
if (best.threshold > 0){
l=k;
u=d->size[best.feature];
first = root->left;
second = root->right;
}
else{
l=0;
u=k;
first = root->right;
second = root->left;
}
/* Here's how this works when threshold > 0. The case where threshold < 0 is analogous:
* Let X be the set of all examples whose feature best.feature has value > threshold
* For every x in X decrease valid[x] by 1.
* This leads to valid[x] > 0 iff x was previously valid and has value < threshold
*
* build left subtree (using the valid examples)
*
* For every x in X increase valid[x] by 2.
* For every example x decrease valid[x] by 1.
* This leads to valid[x] > 0 iff x was previously valid and has value > threshold
*
* build right subtree (using the valid examples)
*
* Finally restore:
* For every x in X decrease valid[x] by 1.
* For every example x increase valid[x] by 1.
* This makes valid obtain its original state
* (One can verify this by adding up all the transformations)
*/
for(i=l; i<u; i++)
t->valid[b[i].example]-=1;
growrec(t, first, d, depth+1);
for(i=l; i<u; i++)
t->valid[b[i].example]+=2;
for(i=0; i<d->nex; i++)
t->valid[i]-=1;
growrec(t, second, d, depth+1);
for(i=l; i<u; i++)
t->valid[b[i].example]-=1;
for(i=0; i<d->nex; i++)
t->valid[i]+=1;
/* Unmark the feature */
if(!d->cont[best.feature])
t->used[best.feature]=0;
}
int
getrec(void)
{
register char *rr;
extern int svargc;
extern char **svargv;
register int c, sep, k, m, n;
wchar_t wc;
dprintf("**RS=%o, **FS=%o\n", **RS, **FS);
donefld = 0;
donerec = 1;
record[0] = 0;
while (svargc > 0) {
dprintf("svargc=%d, *svargv=%s\n", svargc, *svargv);
if (infile == NULL) { /* have to open a new file */
if (member('=', *svargv)) { /* it's a var=value argument */
setclvar(*svargv);
svargv++;
svargc--;
continue;
}
*FILENAME = file = *svargv;
dprintf("opening file %s\n", file);
if (*file == '-') {
if (yyin == stdin && ! lexprog)
error(FATAL, "standard input already used for reading commands");
else
infile = stdin;
}
else if ((infile = fopen(file, "r")) == NULL)
error(FATAL, "can't open %s", file);
}
next(wc, *RS, n);
if ((sep = **RS) == 0)
sep = '\n';
for (rr = record; ; ) {
cont: for (; (c=getc(infile)) != sep && c != EOF; *rr++ = c) {
if (rr >= record+RECSIZE-n-3) {
size_t diff = rr - record;
growrec();
rr = &record[diff];
}
}
if (c != EOF) {
/*
* Note: This code does not restrict occurences
* of the multibyte sequence in RS to the start
* of an input character.
*/
for (m = 1; m < n; m++) {
if ((c = getc(infile)) == EOF ||
c != (*RS)[m]) {
for (k = 0; k < m; k++)
*rr++ = (*RS)[k];
if (c == EOF)
break;
*rr++ = c;
goto cont;
}
}
}
if (**RS == sep || c == EOF)
break;
if ((c = getc(infile)) == '\n' || c == EOF) /* 2 in a row */
break;
if (rr >= record+RECSIZE-n-3) {
size_t diff = rr - record;
growrec();
rr = &record[diff];
}
*rr++ = '\n';
*rr++ = c;
}
*rr = 0;
if (mustfld)
fldbld();
if (c != EOF || rr > record) { /* normal record */
recloc->tval &= ~NUM;
recloc->tval |= STR;
++nrloc->fval;
nrloc->tval &= ~STR;
nrloc->tval |= NUM;
return(1);
}
/* EOF arrived on this file; set up next */
if (infile != stdin)
fclose(infile);
infile = NULL;
svargc--;
svargv++;
}
return(0); /* true end of file */
}