int
main(int argc, char *argv[]) {
static VShort type = 0;
static VArgVector contrast;
static VOptionDescRec options[] = {
{"type", VShortRepn, 1, (VPointer) &type, VOptionalOpt, TYPDict, "type of output"},
{"contrast", VFloatRepn, 0, (VPointer) &contrast, VRequiredOpt, NULL, "contrast vector"}
};
FILE *in_file, *out_file;
VAttrList list = NULL, out_list = NULL;
gsl_vector_float *cont;
float u;
int i;
char prg_name[100];
char ver[100];
getLipsiaVersion(ver, sizeof(ver));
sprintf(prg_name, "vgetcontrast V%s", ver);
fprintf(stderr, "%s\n", prg_name);
VParseFilterCmd(VNumber(options), options, argc, argv, &in_file, &out_file);
cont = gsl_vector_float_alloc(contrast.number);
for(i = 0; i < contrast.number; i++) {
u = ((VFloat *)contrast.vector)[i];
fvset(cont, i, u);
}
if(!(list = VReadFile(in_file, NULL)))
exit(1);
fclose(in_file);
out_list = VGetContrast(list, cont, type);
/* Output: */
VHistory(VNumber(options), options, prg_name, &list, &out_list);
if(! VWriteFile(out_file, out_list))
exit(1);
fprintf(stderr, "%s: done.\n", argv[0]);
return 0;
}
gsl_vector_float * SpikingGroup::get_state_vector(string key)
{
if ( state_vector.find(key) == state_vector.end() ) {
gsl_vector_float * vec = gsl_vector_float_alloc (get_vector_size());
state_vector[key] = vec;
return vec;
} else {
return state_vector.find(key)->second;
}
}
/*
** general linear regression
*/
VAttrList
VRegression(ListInfo *linfo, int nlists, VShort minval, VImage design, VFloat sigma, VLong itr) {
VAttrList out_list;
VImageInfo *xinfo;
int nbands = 0, nslices = 0, nrows = 0, ncols = 0, slice, row, col, nr, nc;
VImage src[NSLICES], res_image = NULL;
VImage beta_image[MBETA], BCOV = NULL, KX_image = NULL;
VImage res_map[ETMP];
float smooth_fwhm = 0, vx = 0, vy = 0, vz = 0;
VFloat *float_pp, df;
VRepnKind repn;
float d, err;
int i, k, l, n, m = 0, nt, fd = 0, npix = 0;
int i0 = 0, i1 = 0;
double u, sig, trace = 0, trace2 = 0, var = 0, sum = 0, nx = 0, mean = 0, sum2;
float *ptr1, *ptr2;
double x;
gsl_matrix_float *X = NULL, *XInv = NULL, *SX = NULL;
gsl_vector_float *y, *z, *beta, *ys;
gsl_vector *kernel;
gsl_matrix_float *S = NULL, *Vc = NULL, *F = NULL, *P = NULL, *Q = NULL;
gsl_matrix_float *R = NULL, *RV = NULL;
VBoolean smooth = TRUE; /* no smoothness estimation */
gsl_set_error_handler_off();
/*
** read input data
*/
nslices = nbands = nrows = ncols = 0;
for(k = 0; k < nlists; k++) {
n = linfo[k].nslices;
nr = linfo[k].nrows;
nc = linfo[k].ncols;
nt = linfo[k].ntimesteps;
nbands += nt;
if(nslices == 0)
nslices = n;
else if(nslices != n)
VError(" inconsistent image dimensions, slices: %d %d", n, nslices);
if(nrows == 0)
nrows = nr;
else if(nrows != nr)
VError(" inconsistent image dimensions, rows: %d %d", nr, nrows);
if(ncols == 0)
ncols = nc;
else if(ncols != nc)
VError(" inconsistent image dimensions, cols: %d %d", nc, ncols);
}
fprintf(stderr, " num images: %d, image dimensions: %d x %d x %d\n",
nlists, nslices, nrows, ncols);
/*
** get design dimensions
*/
m = VImageNRows(design); /* number of timesteps */
n = VImageNColumns(design); /* number of covariates */
fprintf(stderr, " ntimesteps=%d, num covariates=%d\n", m, n);
if(n >= MBETA)
VError(" too many covariates (%d), max is %d", n, MBETA);
if(m != nbands)
VError(" design dimension inconsistency: %d %d", m, nbands);
fprintf(stderr, " working...\n");
/*
** read design matrix
*/
X = gsl_matrix_float_alloc(m, n);
for(k = 0; k < m; k++) {
for(l = 0; l < n; l++) {
x = VGetPixel(design, 0, k, l);
fmset(X, k, l, (float)x);
}
}
/*
** pre-coloring, set up K-matrix, S=K, V = K*K^T with K=S
*/
S = gsl_matrix_float_alloc(m, m);
GaussMatrix((double)sigma, S);
Vc = fmat_x_matT(S, S, NULL);
/*
** compute pseudoinverse
*/
SX = fmat_x_mat(S, X, NULL);
XInv = fmat_PseudoInv(SX, NULL);
/*
** get variance estimate
*/
Q = fmat_x_mat(XInv, Vc, Q);
F = fmat_x_matT(Q, XInv, F);
BCOV = VCreateImage(1, n, n, VFloatRepn);
float_pp = VImageData(BCOV);
ptr1 = F->data;
for(i = 0; i < n * n; i++)
*float_pp++ = *ptr1++;
gsl_matrix_float_free(Q);
gsl_matrix_float_free(F);
/*
** get effective degrees of freedom
*/
R = gsl_matrix_float_alloc(m, m);
P = fmat_x_mat(SX, XInv, P);
gsl_matrix_float_set_identity(R);
gsl_matrix_float_sub(R, P);
RV = fmat_x_mat(R, Vc, NULL);
trace = 0;
for(i = 0; i < m; i++)
trace += fmget(RV, i, i);
P = fmat_x_mat(RV, RV, P);
trace2 = 0;
for(i = 0; i < m; i++)
trace2 += fmget(P, i, i);
df = (trace * trace) / trace2;
fprintf(stderr, " df= %.3f\n", df);
/*
** create output images
*/
xinfo = linfo[0].info;
out_list = VCreateAttrList();
res_image = VCreateImage(nslices, nrows, ncols, VFloatRepn);
VFillImage(res_image, VAllBands, 0);
VSetAttr(VImageAttrList(res_image), "name", NULL, VStringRepn, "RES/trRV");
VSetAttr(VImageAttrList(res_image), "modality", NULL, VStringRepn, "RES/trRV");
VSetAttr(VImageAttrList(res_image), "df", NULL, VFloatRepn, df);
VSetAttr(VImageAttrList(res_image), "patient", NULL, VStringRepn, xinfo->patient);
VSetAttr(VImageAttrList(res_image), "voxel", NULL, VStringRepn, xinfo->voxel);
VSetAttr(VImageAttrList(res_image), "repetition_time", NULL, VLongRepn, itr);
VSetAttr(VImageAttrList(res_image), "talairach", NULL, VStringRepn, xinfo->talairach);
/* neu */
VSetAttr(VImageAttrList(res_image),"indexOrigin",NULL,VStringRepn,xinfo->indexOrigin);
VSetAttr(VImageAttrList(res_image),"columnVec",NULL,VStringRepn,xinfo->columnVec);
VSetAttr(VImageAttrList(res_image),"rowVec",NULL,VStringRepn,xinfo->rowVec);
VSetAttr(VImageAttrList(res_image),"sliceVec",NULL,VStringRepn,xinfo->sliceVec);
VSetAttr(VImageAttrList(res_image),"FOV",NULL,VStringRepn,xinfo->FOV);
/*--------*/
if(xinfo->fixpoint[0] != 'N')
VSetAttr(VImageAttrList(res_image), "fixpoint", NULL, VStringRepn, xinfo->fixpoint);
if(xinfo->ca[0] != 'N') {
VSetAttr(VImageAttrList(res_image), "ca", NULL, VStringRepn, xinfo->ca);
VSetAttr(VImageAttrList(res_image), "cp", NULL, VStringRepn, xinfo->cp);
VSetAttr(VImageAttrList(res_image), "extent", NULL, VStringRepn, xinfo->extent);
}
VAppendAttr(out_list, "image", NULL, VImageRepn, res_image);
for(i = 0; i < n; i++) {
beta_image[i] = VCreateImage(nslices, nrows, ncols, VFloatRepn);
VFillImage(beta_image[i], VAllBands, 0);
VSetAttr(VImageAttrList(beta_image[i]), "patient", NULL, VStringRepn, xinfo->patient);
VSetAttr(VImageAttrList(beta_image[i]), "voxel", NULL, VStringRepn, xinfo->voxel);
VSetAttr(VImageAttrList(beta_image[i]), "repetition_time", NULL, VLongRepn, itr);
VSetAttr(VImageAttrList(beta_image[i]), "talairach", NULL, VStringRepn, xinfo->talairach);
/* neu */
VSetAttr(VImageAttrList(beta_image[i]),"indexOrigin",NULL,VStringRepn,xinfo->indexOrigin);
VSetAttr(VImageAttrList(beta_image[i]),"columnVec",NULL,VStringRepn,xinfo->columnVec);
VSetAttr(VImageAttrList(beta_image[i]),"rowVec",NULL,VStringRepn,xinfo->rowVec);
VSetAttr(VImageAttrList(beta_image[i]),"sliceVec",NULL,VStringRepn,xinfo->sliceVec);
VSetAttr(VImageAttrList(beta_image[i]),"FOV",NULL,VStringRepn,xinfo->FOV);
/*--------*/
if(xinfo->fixpoint[0] != 'N')
VSetAttr(VImageAttrList(beta_image[i]), "fixpoint", NULL, VStringRepn, xinfo->fixpoint);
if(xinfo->ca[0] != 'N') {
VSetAttr(VImageAttrList(beta_image[i]), "ca", NULL, VStringRepn, xinfo->ca);
VSetAttr(VImageAttrList(beta_image[i]), "cp", NULL, VStringRepn, xinfo->cp);
VSetAttr(VImageAttrList(beta_image[i]), "extent", NULL, VStringRepn, xinfo->extent);
}
VSetAttr(VImageAttrList(beta_image[i]), "name", NULL, VStringRepn, "BETA");
VSetAttr(VImageAttrList(beta_image[i]), "modality", NULL, VStringRepn, "BETA");
VSetAttr(VImageAttrList(beta_image[i]), "beta", NULL, VShortRepn, i + 1);
VSetAttr(VImageAttrList(beta_image[i]), "df", NULL, VFloatRepn, df);
VAppendAttr(out_list, "image", NULL, VImageRepn, beta_image[i]);
}
VSetAttr(VImageAttrList(design), "name", NULL, VStringRepn, "X");
VSetAttr(VImageAttrList(design), "modality", NULL, VStringRepn, "X");
VAppendAttr(out_list, "image", NULL, VImageRepn, design);
KX_image = Mat2Vista(SX);
VSetAttr(VImageAttrList(KX_image), "name", NULL, VStringRepn, "KX");
VSetAttr(VImageAttrList(KX_image), "modality", NULL, VStringRepn, "KX");
VAppendAttr(out_list, "image", NULL, VImageRepn, KX_image);
VSetAttr(VImageAttrList(BCOV), "name", NULL, VStringRepn, "BCOV");
VSetAttr(VImageAttrList(BCOV), "modality", NULL, VStringRepn, "BCOV");
VAppendAttr(out_list, "image", NULL, VImageRepn, BCOV);
/*
** create temporary images for smoothness estimation
*/
/* smoothness estim only for 3D images, i.e. CA/CP known */
/* if(xinfo->ca[0] == 'N')
smooth = FALSE;*/
if(smooth) {
i0 = 20;
i1 = i0 + 30;
if(i1 > m)
i1 = m;
for(i = i0; i < i1; i++) {
if(i - i0 >= ETMP)
VError(" too many tmp images");
res_map[i - i0] = VCreateImage(nslices, nrows, ncols, VFloatRepn);
VFillImage(res_map[i - i0], VAllBands, 0);
}
}
/*
** process
*/
ys = gsl_vector_float_alloc(m);
y = gsl_vector_float_alloc(m);
z = gsl_vector_float_alloc(m);
beta = gsl_vector_float_alloc(n);
kernel = GaussKernel((double)sigma);
for(k = 0; k < nlists; k++) {
src[k] = VCreateImage(linfo[k].ntimesteps, nrows, ncols, linfo[k].repn);
VFillImage(src[k], VAllBands, 0);
}
npix = 0;
for(slice = 0; slice < nslices; slice++) {
if(slice % 5 == 0)
fprintf(stderr, " slice: %3d\r", slice);
for(k = 0; k < nlists; k++) {
if(linfo[k].zero[slice] == 0)
goto next1;
fd = open(linfo[k].filename, O_RDONLY);
if(fd == -1)
VError("could not open file %s", linfo[k].filename);
nt = linfo[k].ntimesteps;
if(! VReadBandDataFD(fd, &linfo[k].info[slice], 0, nt, &src[k]))
VError(" error reading data");
close(fd);
}
repn = linfo[0].repn;
for(row = 0; row < nrows; row++) {
for(col = 0; col < ncols; col++) {
for(k = 0; k < nlists; k++)
if(VPixel(src[k], 0, row, col, VShort) < minval + 1)
goto next;
npix++;
/* read time series data */
sum = sum2 = nx = 0;
ptr1 = y->data;
for(k = 0; k < nlists; k++) {
nt = VImageNBands(src[k]);
for(i = 0; i < nt; i++) {
u = VPixel(src[k], i, row, col, VShort);
(*ptr1++) = u;
sum += u;
sum2 += u * u;
nx++;
}
}
mean = sum / nx;
sig = sqrt((double)((sum2 - nx * mean * mean) / (nx - 1.0)));
if(sig < 0.001)
continue;
/* centering and scaling, Seber, p.330 */
ptr1 = y->data;
for(i = 0; i < m; i++) {
u = ((*ptr1) - mean) / sig;
(*ptr1++) = u + 100.0;
}
/* S x y */
ys = VectorConvolve(y, ys, kernel);
/* compute beta's */
fmat_x_vector(XInv, ys, beta);
/* residuals */
fmat_x_vector(SX, beta, z);
err = 0;
ptr1 = ys->data;
ptr2 = z->data;
for(i = 0; i < m; i++) {
d = ((*ptr1++) - (*ptr2++));
err += d * d;
}
/* sigma^2 */
var = err / trace;
/* write residuals output */
VPixel(res_image, slice, row, col, VFloat) = (VFloat)var;
/* save residuals of several timesteps for smoothness estimation */
if(smooth) {
ptr1 = ys->data;
ptr2 = z->data;
err = 0;
for(i = i0; i < i1; i++) {
d = ((*ptr1++) - (*ptr2++));
err += d * d;
VPixel(res_map[i - i0], slice, row, col, VFloat) = d;
}
if (err > 0) err = sqrt(err);
for(i = i0; i < i1; i++) {
d = VPixel(res_map[i - i0], slice, row, col, VFloat);
if (err > 1.0e-6) d /= err;
else d = 0;
VPixel(res_map[i - i0], slice, row, col, VFloat) = d / err;
}
}
/* write beta output */
ptr1 = beta->data;
for(i = 0; i < n; i++)
VPixel(beta_image[i], slice, row, col, VFloat) = (VFloat)(*ptr1++);
next:
;
}
}
next1:
;
}
/*
** Smoothness estimation based on residual images
*/
if(smooth) {
smooth_fwhm = VSmoothnessEstim(res_map, i1 - i0);
sscanf(xinfo->voxel, "%f %f %f", &vx, &vy, &vz);
vx = (vx + vy + vz) / 3.0; /* voxels should be isotropic */
smooth_fwhm *= vx;
fprintf(stderr, " smoothness: %f\n", smooth_fwhm);
VSetAttr(VImageAttrList(res_image), "smoothness", NULL, VFloatRepn, smooth_fwhm);
for(i = 0; i < n; i++) {
VSetAttr(VImageAttrList(beta_image[i]), "smoothness", NULL, VFloatRepn, smooth_fwhm);
}
for(i = 0; i < i1 - i0; i++)
VDestroyImage(res_map[i]);
}
ende:
if(npix == 0)
VError(" no voxels above threshold %d found", minval);
return out_list;
}
VAttrList
VGetContrast(VAttrList list, gsl_vector_float *con, VShort type) {
VAttrList out_list;
int nbands = 0, nrows = 0, ncols = 0, band, row, col;
VImage src = NULL, dest = NULL, std_image = NULL;
VImage beta_images[MBETA], res_image = NULL, bcov_image = NULL;
VString buf = NULL;
VAttrListPosn posn;
VString str;
int i, nbeta;
float t = 0, s = 0, tsigma = 0, z = 0, zmax = 0, zmin = 0;
float sigma, var, sum, df;
float *ptr1, *ptr2;
char *constring = NULL;
gsl_vector_float *beta = NULL, *tmp = NULL;
gsl_matrix_float *bcov = NULL;
i = 0;
for(VFirstAttr(list, & posn); VAttrExists(& posn); VNextAttr(& posn)) {
if(VGetAttrRepn(& posn) != VImageRepn)
continue;
VGetAttrValue(& posn, NULL, VImageRepn, & src);
if(VPixelRepn(src) != VFloatRepn)
continue;
VGetAttr(VImageAttrList(src), "modality", NULL, VStringRepn, &str);
if(strcmp(str, "BETA") == 0) {
beta_images[i++] = VCopyImage(src, NULL, VAllBands);
} else if(strcmp(str, "RES/trRV") == 0) {
res_image = VCopyImage(src, NULL, VAllBands);
} else if(strcmp(str, "BCOV") == 0) {
bcov_image = VCopyImage(src, NULL, VAllBands);
}
}
nbeta = VImageNRows(bcov_image);
nbands = VImageNBands(beta_images[0]);
nrows = VImageNRows(beta_images[0]);
ncols = VImageNColumns(beta_images[0]);
if(VGetAttr(VImageAttrList(beta_images[0]), "df", NULL, VFloatRepn, &df) != VAttrFound)
VError(" attribute 'df' not found");
if(nbeta > MBETA) {
fprintf(stderr, " number of betas: %d, maximum number of betas: %d\n", nbeta, MBETA);
VError(" maximum number of betas is exceeded");
}
/*
** read contrast vector
*/
if(nbeta != con->size)
VError("contrast vector has bad length (%d), correct length is %d", con->size, nbeta);
fprintf(stderr, " contrast vector:\n");
char str1[10];
constring = (char *)VMalloc(sizeof(char) * 10 * nbeta);
constring[0] = '\0';
for(i = 0; i < nbeta; i++) {
fprintf(stderr, " %.2f", fvget(con, i));
sprintf(str1, "%1.2f ", fvget(con, i));
strcat((char *)constring, (const char *)str1);
}
fprintf(stderr, "\n");
/* get variance estimation */
bcov = gsl_matrix_float_alloc(nbeta, nbeta);
ptr1 = VImageData(bcov_image);
ptr2 = bcov->data;
for(i = 0; i < nbeta * nbeta; i++)
*ptr2++ = *ptr1++;
gsl_matrix_float_transpose(bcov);
tmp = fmat_x_vector(bcov, con, tmp);
var = fskalarproduct(tmp, con);
sigma = sqrt(var);
/*
** create output data structs
*/
out_list = VCreateAttrList();
dest = VCreateImage(nbands, nrows, ncols, VFloatRepn);
VFillImage(dest, VAllBands, 0);
VCopyImageAttrs(beta_images[0], dest);
switch(type) {
case 0: /* conimg */
buf = VNewString("conimg");
break;
case 1: /* t-image */
buf = VNewString("tmap");
break;
case 2: /* zmap */
buf = VNewString("zmap");
break;
default:
VError(" illegal type");
}
fprintf(stderr, " output type: %s\n", buf);
VSetAttr(VImageAttrList(dest), "modality", NULL, VStringRepn, buf);
VSetAttr(VImageAttrList(dest), "name", NULL, VStringRepn, buf);
VSetAttr(VImageAttrList(dest), "contrast", NULL, VStringRepn, constring);
VAppendAttr(out_list, "image", NULL, VImageRepn, dest);
if(type == 0) {
std_image = VCreateImage(nbands, nrows, ncols, VFloatRepn);
VFillImage(std_image, VAllBands, 0);
VCopyImageAttrs(beta_images[0], std_image);
VSetAttr(VImageAttrList(std_image), "modality", NULL, VStringRepn, "std_dev");
VSetAttr(VImageAttrList(std_image), "name", NULL, VStringRepn, "std_dev");
VAppendAttr(out_list, "image", NULL, VImageRepn, std_image);
}
/*
** loop thru image
*/
zmax = zmin = 0;
beta = gsl_vector_float_alloc(nbeta);
for(band = 0; band < nbands; band++) {
for(row = 0; row < nrows; row++) {
for(col = 0; col < ncols; col++) {
t = z = sum = 0;
ptr1 = beta->data;
for(i = 0; i < nbeta; i++) {
*ptr1++ = VPixel(beta_images[i], band, row, col, VFloat);
}
sum = fskalarproduct(beta, con);
if(ABS(sum) < 1.0e-10)
continue;
s = VPixel(res_image, band, row, col, VFloat);
tsigma = sqrt(s) * sigma;
if(tsigma > 0.00001)
t = sum / tsigma;
else
t = 0;
if(isnan(t) || isinf(t))
t = 0;
switch(type) {
case 0: /* conimg */
z = sum;
break;
case 1: /* t-image */
z = t;
break;
case 2: /* zmap */
z = t2z_approx(t, df);
if(z > 30)
z = 30;
if(sum < 0)
z = -z;
break;
default:
;
}
if(isnan(z) || isinf(z))
z = 0;
if(z > zmax)
zmax = z;
if(z < zmin)
zmin = z;
VPixel(dest, band, row, col, VFloat) = z;
if(type == 0)
VPixel(std_image, band, row, col, VFloat) = tsigma;
}
}
}
fprintf(stderr, " min= %.3f, max= %.3f\n", zmin, zmax);
return out_list;
}
ContrastiveDivergence::ContrastiveDivergence(float momentum, int k, int batchsize, int e) : momentum(momentum), k(k), batchsize(batchsize), epochs(e)
{
monitor = NULL;
identity = gsl_vector_float_alloc(batchsize);
gsl_vector_float_set_all(identity, 1);
}
