void jitter::set(const idx<t_jitter> &j) {
s = j.get(0);
h = (int) j.get(1);
w = (int) j.get(2);
r = j.get(3);
idx_copy(j, jitts);
}
unsigned int draw_layer(idx<float> &img, int dimn, int layern, const char *s,
unsigned int h, unsigned int w) {
#ifdef __GUI__
idx<float> layer = img.select(dimn, layern);
draw_matrix(layer, s, h, w, 1.0, 1.0, (float)-1.0, (float)1.0);
#endif
return img.dim(1) + 3;
}
void matlab::read_cast_matrix(mxArray *var, idx<T> &m) {
#ifdef __MATLAB__
// allocate a temporary matrix with same type as original matrix type
idx<Tmatlab> tmp(m.get_idxdim());
// load data
void *data = mxGetData(var);
// copy to idx
memcpy(m.idx_ptr(), (Tmatlab*) data, m.nelements() * sizeof (Tmatlab));
// copy-cast
idx_copy(tmp, m);
#endif
}
void padder<T>::pad(idx<T> &in, idx<T> &out) {
// allocate padded buffer
idxdim d = in;
d.setdim(1, d.dim(1) + nrow + nrow2);
d.setdim(2, d.dim(2) + ncol + ncol2);
if (out.get_idxdim() != d)
out.resize(d);
idx_clear(out);
// copy in to padded buffer
idx<T> tmp = out.narrow(1, in.dim(1), nrow);
tmp = tmp.narrow(2, in.dim(2), ncol);
idx_copy(in, tmp);
if (bmirror)
mirror(in, out);
}
idx<T> padder<T>::pad(idx<T> &in) {
// allocate padded buffer
idxdim d = in;
d.setdim(1, d.dim(1) + nrow + nrow2);
d.setdim(2, d.dim(2) + ncol + ncol2);
idx<T> out(d);
idx_clear(out);
// copy in to padded buffer
idx<T> tmp = out.narrow(1, in.dim(1), nrow);
tmp = tmp.narrow(2, in.dim(2), ncol);
idx_copy(in, tmp);
if (bmirror)
mirror(in, out);
// return result
return out;
}
unsigned int draw_layer(idx<float> &layer, const char *s,
unsigned int h, unsigned int w) {
#ifdef __GUI__
draw_matrix(layer, s, h, w, 1.0, 1.0, (float)-1.0, (float)1.0);
#endif
return layer.dim(1) + 3;
}
// display original and new images.
void display_images(idx<ubyte> &classes, int label, idx<float> &original,
idx<float> &new_images, const int channels_mode,
idx<ubyte> &ds_names, int current_ds) {
#ifdef __GUI__
unsigned int h = 0, w = 0;
static unsigned int cnt = 0;
idx<float> layer;
// only display every 15 images
if (cnt++ % 15 != 0) return ;
// reset window and do a batch display
disable_window_updates();
clear_window();
// original image (RGB)
static string s;
s = "RGB"; s += " - "; s += (char *) classes[label].idx_ptr();
draw_matrix(original, s.c_str(), h, w);
w = 0; h += new_images.dim(1) + 5;
gui << at(h, w) << black_on_white();
gui << ds_names[current_ds].idx_ptr() << " dataset:";
h += 15;
// new images
idx_bloop1(image, new_images, float) {
switch (channels_mode) {
case 1: // YpUV
w += draw_layer(image, 2, 0, "Yp", h, w);
w += draw_layer(image, 2, 1, "U", h, w);
w += draw_layer(image, 2, 2, "V", h, w);
break ;
case 3: // Yp
w += draw_layer(image, 2, 0, "Yp", h, w);
break ;
case 4: { // YpH3
w += draw_layer(image, 2, 0, "Yp", h, w);
w += draw_layer(image, 2, 1, "H3", h, w);
idx_addc(layer, (float)1.0, layer);
idx_dotc(layer, (float)210.0, layer);
w += layer.dim(1) + 5;
static idx<float> rgb(layer.dim(0), layer.dim(1), 3);
h3_to_rgb(layer, rgb);
draw_matrix(rgb, "H3 (colored)", h, w);
break ;
}
case 5: // VpH2SV
w += draw_layer(image, 2, 0, "Vp", h, w);
w += draw_layer(image, 2, 1, "H1", h, w);
w += draw_layer(image, 2, 2, "H2", h, w);
w += draw_layer(image, 2, 3, "S", h, w);
w += draw_layer(image, 2, 4, "V", h, w);
w = 0; h += image.dim(1) + 5;
break ;
}
}
enable_window_updates();
#endif
}
idx<T> image_region_to_rect(idx<T> &im, const rect<int> &r, uint oheight,
uint owidth, rect<int> *cropped,
uint dh, uint dw)
{
// TODO: check that rectangle is within image
if (im.order() != 2 && im.order() != 3)
eblerror("expected a 2d or 3d input but got " << im);
idxdim d(im);
d.setdim(dh, oheight);
d.setdim(dw, owidth);
idx<T> res(d);
float hcenter = r.h0 + (float)r.height / 2; // input height center
float wcenter = r.w0 + (float)r.width / 2; // input width center
// // limit centers to half the width/height away from borders
// // to handle incorrect regions
// hcenter = MIN((float)im.dim(dh) - (float)r.height/2,
// std::max((float)r.height/2, hcenter));
// wcenter = MIN((float)im.dim(dw) - (float)r.width/2,
// std::max((float)r.width/2, wcenter));
float h0 = hcenter - (float)oheight / 2; // out height offset in input
float w0 = wcenter - (float)owidth / 2; // out width offset in input
float h1 = hcenter + (float)oheight / 2;
float w1 = wcenter + (float)owidth / 2;
int gh0 = (int)std::max(0, (int)MIN(im.dim(dh), h0)); // input h offset
int gw0 = (int)std::max(0, (int)MIN(im.dim(dw), w0)); // input w offset
int gh1 = (int)std::max(0, (int)MIN(im.dim(dh), h1));
int gw1 = (int)std::max(0, (int)MIN(im.dim(dw), w1));
int h = gh1 - gh0 + std::max(0, -r.h0); // out height narrow
int w = gw1 - gw0 + std::max(0, -r.w0); // out width narrow
int fh0 = (int)std::max(0, (int)(gh0 - h0)); // out height offset narrow
int fw0 = (int)std::max(0, (int)(gw0 - w0)); // out width offset narrow
// narrow original image
int hmin = std::max(0, std::min((int)res.dim(dh) - fh0,
std::min((int)im.dim(dh) - gh0, h)));
int wmin = std::max(0, std::min((int)res.dim(dw) - fw0,
std::min((int)im.dim(dw) - gw0, w)));
// clear result image
idx_clear(res);
if (hmin != 0 && wmin != 0) // only copy if overlap with image
{
idx<T> tmpim = im.narrow(dh, hmin, gh0);
tmpim = tmpim.narrow(dw, wmin, gw0);
// narrow target image
idx<T> tmpres = res.narrow(dh, hmin, fh0);
tmpres = tmpres.narrow(dw, wmin, fw0);
// copy original to target
idx_copy(tmpim, tmpres);
}
// set cropped rectangle to region in the output image containing input
if (cropped)
{
cropped->h0 = fh0;
cropped->w0 = fw0;
cropped->height = hmin;
cropped->width = wmin;
}
return res;
}
void serialize(Archive & ar, idx<ubyte>& mat, const unsigned int version) {
intg d1, d2, d3;
if (Archive::is_saving::value) { // saving to stream
if (!mat.contiguousp())
eblerror("expected contiguous idx for serialization");
if (mat.order() != 3)
eblerror("no support for idx order != 3 for now, got: " << mat);
d1 = mat.dim(0);
d2 = mat.dim(1);
d3 = mat.dim(2);
ar & d1;
ar & d2;
ar & d3;
idx_aloop1(m, mat, ubyte) {
ar & *m;
}
} else { // loading from stream
idxlooper<T>::idxlooper(idx<T> &m, int ld) : idx<T>((dummyt*)0) {
if (m.order() == 0) // TODO: allow looping once on 0-order idx
eblerror("cannot loop on idx with order 0. idx is: " << m);
i = 0;
dimd = m.spec.dim[ld];
modd = m.spec.mod[ld];
m.spec.select_into(&(this->spec), ld, i);
this->storage = m.storage;
this->storage->lock();
}
bool detection_thread<T>::set_data(idx<ubyte> &frame2,
std::string &fullname,
std::string &name, uint id) {
// lock data (non blocking)
if (!mutex_in.trylock())
return false;
// check frame is correctly allocated, if not, allocate.
if (frame2.order() != uframe.order())
uframe = idx<ubyte>(frame2.get_idxdim());
else if (frame2.get_idxdim() != uframe.get_idxdim())
uframe.resize(frame2.get_idxdim());
idx_copy(frame2, uframe); // copy frame
frame_loaded = true; // frame is loaded
frame_fullname = fullname;
frame_name = name; // copy name
frame_id = id; // copy frame_id
in_updated = true; // reset updated flag
bavailable = false; // declare thread as not available
bfed = true; // data has been fed at least once
mutex_in.unlock(); // unlock data
return true; // confirm that we copied data.
}
void image_paste_center(idx<T> &in, idx<T> &out) {
if (in.order() != 3 || out.order() != 3)
eblerror("expected 3d idx but got " << in << " and " << out);
int d0 = 1;
int d1 = d0 + 1;
intg ci = in.dim(d0) - out.dim(d0);
intg cj = in.dim(d1) - out.dim(d1);
intg xci = (int) (ci / 2);
intg xcj = (int) (cj / 2);
idx<T> i = in.narrow(d0, out.dim(d0), xci);
i = i.narrow(d1, out.dim(d1), xcj);
idx_copy(i, out);
}
//! Return an idx of dimensions Nx2 containing all possible N similar pairs.
idx<int> make_pairs(idx<int> &labels) {
// allocate maximum number of pairs
idx<int> pairs((labels.dim(0) * (labels.dim(0) - 1)) / 2, 2);
int n = 0, r = 0;
// for each label, loop over all following labels to find pairs
for (int i = 0; i < labels.dim(0) - 1; ++i) {
for (int j = i + 1; j < labels.dim(0); ++j) {
if (labels.get(i) == labels.get(j)) {
r = drand(0.0, 1.0); // randomize distribution as 1st or 2nd element
pairs.set(i, n, (r > 0.5) ? 0 : 1);
pairs.set(j, n, (r > 0.5) ? 1 : 0);
n++;
}
}
}
pairs.resize(n, pairs.dim(1));
return pairs;
}
bool tracking_thread<Tnet>::get_data(vector<bbox*> &bboxes2,
idx<ubyte> &frame2, idx<ubyte> &tpl2) {
// lock data
pthread_mutex_lock(&mutex_out);
// only read data if it has been updated
if (!out_updated) {
// unlock data
pthread_mutex_unlock(&mutex_out);
return false;
}
// clear bboxes
for (ibox = bboxes2.begin(); ibox != bboxes2.end(); ++ibox) {
if (*ibox)
delete *ibox;
}
bboxes2.clear();
// copy bboxes pointers (now responsible for deleting them).
for (ibox = bboxes.begin(); ibox != bboxes.end(); ++ibox) {
bboxes2.push_back(*ibox);
}
// check frame is correctly allocated, if not, allocate.
if (frame2.order() != frame.order())
frame2 = idx<ubyte>(frame.get_idxdim());
else if (frame2.get_idxdim() != frame.get_idxdim())
frame2.resize(frame.get_idxdim());
// copy frame
idx_copy(frame, frame2);
// check tpl is correctly allocated, if not, allocate.
if (tpl2.order() != tpl.order())
tpl2 = idx<ubyte>(tpl.get_idxdim());
else if (tpl2.get_idxdim() != tpl.get_idxdim())
tpl2.resize(tpl.get_idxdim());
// copy tpl
idx_copy(tpl, tpl2);
// reset updated flag
out_updated = false;
// unlock data
pthread_mutex_unlock(&mutex_out);
// confirm that we copied data.
return true;
}
bool detection_thread<T>::get_data(bboxes &bboxes2, idx<ubyte> &frame2,
uint &total_saved_, std::string &frame_name_,
uint *id, svector<midx<T> > *samples,
bboxes *bbsamples, bool *skipped) {
// lock data
mutex_out.lock();
// only read data if it has been updated
if (!out_updated) {
// unlock data
mutex_out.unlock();
return false;
}
// data is updated, but just to tell we skipped the frame
if (frame_skipped) {
if (skipped) *skipped = true;
frame_skipped = false;
// reset updated flag
out_updated = false;
// declare thread as available
bavailable = true;
// unlock data
mutex_out.unlock();
return false;
}
if (skipped) *skipped = false;
// clear bboxes
bboxes2.clear();
bboxes2.push_back_new(bbs);
bbs.clear(); // no use for local bounding boxes anymore, clear them
// check frame is correctly allocated, if not, allocate.
if (frame2.order() != uframe.order())
frame2 = idx<ubyte>(uframe.get_idxdim());
else if (frame2.get_idxdim() != uframe.get_idxdim())
frame2.resize(uframe.get_idxdim());
// copy frame
idx_copy(uframe, frame2);
// set total of boxes saved
total_saved_ = total_saved;
// set frame name
frame_name_ = frame_name;
// set frame id
if (id) *id = frame_id;
// overwrite samples
if (samples) {
samples->clear();
samples->push_back_new(returned_samples);
returned_samples.clear();
}
if (bbsamples) {
bbsamples->clear();
bbsamples->push_back_new(returned_samples_bboxes);
returned_samples_bboxes.clear();
}
// reset updated flag
out_updated = false;
// declare thread as available
bavailable = true;
// unlock data
mutex_out.unlock();
// confirm that we copied data.
return true;
}
void read_cast_matrix(FILE *fp, idx<T2> &out) {
idx<T> m(out.get_idxdim());
read_matrix_body(fp, m);
idx_copy(m, out);
}
void padder<T>::mirror(idx<T> &in, idx<T> &padded)
{
idx<T> tmp, tmp2;
int i;
// mirror border left
for (i = std::max(0, (int)(ncol - in.dim(1) / 2)); i < ncol; ++i)
{
tmp2 = in.narrow(1, 1, ncol - i - 1);
tmp = padded.narrow(1, 1, i);
tmp = tmp.narrow(2, in.dim(2), ncol);
idx_copy(tmp2, tmp);
}
// mirror border right
for (i = std::max(0, (int)(ncol - in.dim(1) / 2)); i < ncol; ++i)
{
tmp2 = in.narrow(1, 1, in.dim(1) - ncol - 1 + i);
tmp = padded.narrow(1, 1, padded.dim(1) - 1 - i);
tmp = tmp.narrow(2, in.dim(2), ncol);
idx_copy(tmp2, tmp);
}
// mirror border top using out as input
for (i = std::max(0, (int)(nrow - in.dim(2) / 2)); i < nrow; ++i)
{
tmp2 = padded.narrow(2, 1, nrow + nrow - i - 1);
tmp = padded.narrow(2, 1, i);
idx_copy(tmp2, tmp);
}
// mirror border bottom using out as input
for (i = std::max(0, (int)(nrow - in.dim(2) / 2)); i < nrow; ++i)
{
tmp2 = padded.narrow(2, 1, padded.dim(2) - nrow * 2 - 1 + i);
tmp = padded.narrow(2, 1, padded.dim(2) - 1 - i);
idx_copy(tmp2, tmp);
}
}
cost_module<T1,T2,Tstate1,Tstate2>::cost_module(idx<T1> &targets_)
: targets(targets_), in2(targets.select(0, 0)), energies(targets_.dim(0)) {
}
// convert the input image <src> into the <channels_mode> format.
void convert_image(idx<float> &src, idx<float> &dst, const int channels_mode,
unsigned int fkernel_size) {
idxdim d = idxdim(src), d2 = idxdim(src);
d2.setdim(2, dst.dim(2));
idx<float> in, out, src2(d), tmp(d2), tmp2;
if (fkernel_size > 0) {
src2 = src.narrow(0, src.dim(0) - fkernel_size + 1,
floor(fkernel_size / 2));
src2 = src2.narrow(1, src.dim(1) - fkernel_size + 1,
floor(fkernel_size / 2));
tmp2 = tmp.narrow(0, tmp.dim(0) - fkernel_size + 1,
floor(fkernel_size / 2));
tmp2 = tmp2.narrow(1, tmp.dim(1) - fkernel_size + 1,
floor(fkernel_size / 2));
}
idx_clear(dst);
// narrow dst to fit src if smaller
if (src2.dim(0) < dst.dim(0))
dst = dst.narrow(0, src2.dim(0), floor((dst.dim(0) - src2.dim(0)) / 2));
if (src2.dim(1) < dst.dim(1))
dst = dst.narrow(1, src2.dim(1), floor((dst.dim(1) - src2.dim(1)) / 2));
// converting
switch (channels_mode) {
case 0: // RGB
idx_copy(src2, dst);
image_global_normalization(dst);
break ;
case 1: // YUV
rgb_to_yuv(src, tmp);
in = tmp.select(2, 0);
d = idxdim(in);
out = idx<float>(d);
image_mexican_filter(in, out, 6, fkernel_size);
idx_copy(out, in);
idx_copy(tmp2, dst);
image_global_normalization(dst);
break ;
case 3: { // Y only
rgb_to_y(src, tmp);
in = tmp.select(2, 0);
d = idxdim(in);
out = idx<float>(d);
image_global_normalization(in);
image_local_normalization(in, out, fkernel_size);
idx_copy(out, in);
idx_copy(tmp2, dst);
}
break ;
case 4: // YH3
rgb_to_yh3(src, tmp);
in = tmp.select(2, 0);
d = idxdim(in);
out = idx<float>(d);
image_mexican_filter(in, out, 6, fkernel_size);
image_global_normalization(out);
idx_copy(out, in);
idx_copy(tmp2, dst);
break ;
case 5: // VpH2SV
rgb_to_vph2sv(src, tmp, 6, fkernel_size);
idx_copy(tmp2, dst);
break ;
default:
cerr << "unknown channel mode: " << channels_mode << endl;
eblerror("unknown channel mode");
}
}
int display(list<string>::iterator &ifname,
bool signd, bool load, list<string> *mats) {
//cout << "displaying " << ifname->c_str() << endl;
// conf mode
if (conf)
return display_net<T>(ifname, signd, load, mats);
// mat mode
if (is_matrix(ifname->c_str())) {
idxdim d = get_matrix_dims(ifname->c_str());
if (interleaved)
d.shift_dim(0, 2);
if (save_individually || print
|| !(d.order() == 2 || (d.order() == 3 &&
(d.dim(2) == 1 || d.dim(2) == 3)))) {
// this is probably not an image, just display info and print matrix
string type;
get_matrix_type(ifname->c_str(), type);
idx<T> m = load_matrix<T>(ifname->c_str());
cout << "Matrix " << ifname->c_str() << " is of type " << type
<< " with dimensions " << d << " (min " << idx_min(m)
<< ", max " << idx_max(m) << ", mean " << idx_mean(m)
<< "):" << endl;
m.print();
if (has_multiple_matrices(ifname->c_str())) {
midx<T> ms = load_matrices<T>(ifname->c_str(), true);
// saving sub-matrices
if (save_individually) {
cout << "Saving each sub-matrix of " << *ifname << " individually..."
<< endl;
save_matrices_individually(ms, *ifname, true);
}
// printing sub-matrices
cout << "This file contains " << m << " matrices: ";
if (ms.order() == 1) {
for (intg i = 0; i < ms.dim(0); ++i) {
idx<T> tmp = ms.mget(i);
cout << tmp.info() << " ";
}
} else if (ms.order() == 2) {
for (intg i = 0; i < ms.dim(0); ++i) {
for (intg j = 0; j < ms.dim(1); ++j) {
idx<T> tmp = ms.mget(i, j);
cout << tmp.info() << " ";
}
cout << endl;
}
}
cout << endl;
} else
cout << "This is a single-matrix file." << endl;
return 0;
}
}
// image mode
int loaded = 0;
static idx<T> mat;
uint h = 0, w = 0, rowh = 0, maxh = 0;
list<string>::iterator fname = ifname;
#ifdef __GUI__
disable_window_updates();
clear_window();
if (show_filename) {
gui << at(h, w) << black_on_white() << ifname->c_str();
h += 16;
}
#endif
maxh = h;
for (uint i = 0; i < nh; ++i) {
rowh = maxh;
for (uint j = 0; j < nw; ++j) {
if (fname == mats->end())
fname = mats->begin();
try {
// if (load)
mat = load_image<T>(*fname);
if (print)
cout << *fname << ": " << mat << endl << mat.str() << endl;
// show only some channels
if (chans >= 0)
mat = mat.select(2, chans);
loaded++;
maxh = (std::max)(maxh, (uint) (rowh + mat.dim(0)));
T min = 0, max = 0;
#ifdef __GUI__
if (autorange || signd) {
if (autorange) {
min = idx_min(mat);
max = idx_max(mat);
} else if (signd) {
T matmin = idx_min(mat);
if ((double)matmin < 0) {
min = -1;
max = -1;
}
}
draw_matrix(mat, rowh, w, zoom, zoom, min, max);
} else
draw_matrix(mat, rowh, w, zoom, zoom, (T) range[0], (T) range[1]);
#endif
w += mat.dim(1) + 1;
} catch(string &err) {
ERROR_MSG(err.c_str());
}
fname++;
if (fname == ifname)
break ;
}
if (fname == ifname)
break ;
maxh++;
w = 0;
}
#ifdef __GUI__
// info
if (show_info) {
set_text_colors(0, 0, 0, 255, 255, 255, 255, 200);
gui << mat;
gui << at(15, 0) << *fname;
gui << at(29, 0) << "min: " << idx_min(mat) << " max: " << idx_max(mat);
}
// help
if (show_help) {
h = 0;
w = 0;
uint hstep = 14;
set_text_colors(0, 0, 255, 255, 255, 255, 255, 200);
gui << at(h, w) << "Controls:"; h += hstep;
set_text_colors(0, 0, 0, 255, 255, 255, 255, 200);
gui << at(h, w) << "Right/Space: next image"; h += hstep;
gui << at(h, w) << "Left/Backspace: previous image"; h += hstep;
gui << at(h, w) << "i: image info"; h += hstep;
gui << at(h, w) << "a: auto-range (use min and max as range)"; h += hstep;
gui << at(h, w) << "x/z: show more/less images on width axis"; h += hstep;
gui << at(h, w) << "y/t: show more/less images on height axis"; h += hstep;
gui << at(h, w) << "0,1,2: show channel 0, 1 or 2 only"; h += hstep;
gui << at(h, w) << "9: show alls channels"; h += hstep;
gui << at(h, w) << "h: help";
}
// update title
string title;
title << "matshow: " << ebl::basename(ifname->c_str());
set_window_title(title.c_str());
enable_window_updates();
#endif
return loaded;
}
generic_conv_net(parameter &trainableParam,
intg output_size)
: layers_n<state_idx>(true) { // owns modules
cout << "Initializing ConvNet..." << endl;
//! Define the number of feature maps per layer (C0, C1, C2)
intg featureMaps0 = 6;
intg featureMaps1 = 12;
intg featureMaps2 = 40;
//! Define tables of connections between layers.
//! These two are fully connected layer, i.e. each feature map in a layer
//! is connected to every feature map in the previous layer
table0 = full_table(1, featureMaps0); //! from input to C0
table2 = full_table(featureMaps1, featureMaps2); //! from S1 to C2
//! ... whereas the connections there are sparse (S0 to C1):
table1 = idx<intg>(44, 2); //! from S0 to C1
intg tbl[44][2] =
{{0, 0}, {1, 0}, {2, 0}, //! 0,1,2 in S0 connected to 0 in C1
{1, 1}, {2, 1}, {3, 1}, //! and so on...
{2, 2}, {3, 2}, {4, 2},
{3, 3}, {4, 3}, {5, 3},
{4, 4}, {5, 4}, {0, 4},
{5, 5}, {0, 5}, {1, 5},
{0, 6}, {1, 6}, {2, 6}, {3, 6},
{1, 7}, {2, 7}, {3, 7}, {4, 7},
{2, 8}, {3, 8}, {4, 8}, {5, 8},
{3, 9}, {4, 9}, {5, 9}, {0, 9},
{4, 10}, {5, 10}, {0, 10}, {1, 10},
{0, 11}, {1, 11}, {2, 11}, {3, 11}, {4, 11}, {5, 11}};
memcpy(table1.idx_ptr(), tbl, table1.nelements() * sizeof (intg));
//! Finally we initialize the architecture of the ConvNet.
//! In this case we create a CSCSCF network.
//! It's easy to change the architecture, by simply removing/adding a call
//! to addModule(...)
//! C0 Layer
add_module(new nn_layer_convolution(trainableParam, //! Shared weights
7, 7, //! Dim of kernel
1, 1, //! size of subsampling
table0), //! Conx btwn input layer and C0
//! state_idx holds the feature maps of C0
new state_idx(featureMaps0,1,1));
//! S0 Layer
add_module(new nn_layer_subsampling(trainableParam,
2, 2, //! Dim of stride
2, 2, //! Dim of subsampling mask
featureMaps0),
new state_idx(featureMaps0,1,1));
//! C1 Layer
add_module(new nn_layer_convolution(trainableParam,
7, 7,
1, 1,
table1),
new state_idx(featureMaps1,1,1));
//! S1 Layer
add_module(new nn_layer_subsampling(trainableParam,
2, 2,
2, 2,
featureMaps1),
new state_idx(featureMaps1,1,1));
//! C2 Layer
add_module(new nn_layer_convolution(trainableParam,
7, 7,
1, 1,
table2),
new state_idx(featureMaps2,1,1));
//! F Layer
add_last_module(new nn_layer_full(trainableParam,
featureMaps2,
output_size));
}
//! Recursively goes through dir, looking for files matching extension ext.
void process_dir(const char *dir, const char *ext, const char* leftp,
const char *rightp, unsigned int width,
unsigned int fwidth, idx<float> &images,
idx<int> &labels, int label, bool silent, bool display,
bool *binocular,
const int channels_mode, const int channels_size,
idx<ubyte> &classes, unsigned int &counter,
idx<unsigned int> &counters_used, idx<int> &ds_assignment,
unsigned int fkernel_size, int deformations,
idx<int> &deformid, int &ndefid,
idx<ubyte> &ds_names) {
regex eExt(ext);
string el(".*");
idx<float> limg(1, 1, 1);
idx<float> rimg(1, 1, 1);
idx<float> tmp, left_images;
idx<int> current_labels;
idx<int> current_deformid;
idx<float> tmp2;
int current_ds;
unsigned int current_used;
if (leftp) {
el += leftp;
el += ".*";
}
regex eLeft(el);
cmatch what;
path p(dir);
if (!exists(p))
return ;
directory_iterator end_itr; // default construction yields past-the-end
for (directory_iterator itr(p); itr != end_itr; ++itr) {
if (is_directory(itr->status())) {
process_dir(itr->path().string().c_str(), ext, leftp, rightp, width,
fwidth, images, labels, label, silent, display, binocular,
channels_mode, channels_size, classes, counter,
counters_used, ds_assignment, fkernel_size, deformations,
deformid, ndefid, ds_names);
} else if (regex_match(itr->leaf().c_str(), what, eExt)) {
if (regex_match(itr->leaf().c_str(), what, eLeft)) {
current_ds = ds_assignment.get(counter);
if (current_ds != -1) {
current_used = counters_used.get(current_ds);
// found left image
// increase example number
labels.set(label, current_ds, counters_used.get(current_ds));
// check for right image
if (rightp != NULL) {
regex reg(leftp);
string rp(rightp);
string s = regex_replace(itr->leaf(), reg, rp);
string sfull = itr->path().branch_path().string();
sfull += "/";
sfull += s;
path r(sfull);
if (exists(r)) {
// found right image
*binocular = true;
if (image_read_rgbx(r.string().c_str(), rimg)) {
// resize stereo dimension to twice the channels_size
if (images.dim(3) == channels_size)
images.resize(images.dim(0), images.dim(1),
images.dim(2), images.dim(3) * 2);
// take the right most square of the image
if (rimg.dim(0) <= rimg.dim(1))
rimg = rimg.narrow(1, rimg.dim(0),
rimg.dim(1) - rimg.dim(0));
else
rimg = rimg.narrow(0, rimg.dim(1),
rimg.dim(0) - rimg.dim(1));
// resize image to target width
rimg = image_resize(rimg, width, width, 1);
tmp = images[current_ds];
tmp = tmp[counters_used.get(current_ds)];
tmp = tmp.narrow(2, channels_size, channels_size);
// finally copy right images to idx
convert_image(rimg, tmp, channels_mode, fkernel_size);
}
if (!silent)
cout << "Processing (right): " << sfull << endl;
}
}
if (!silent) {
cout << counter << "/" << ds_assignment.dim(0) << ": ";
cout << itr->path().string().c_str() << endl;
}
// process left image
if (image_read_rgbx(itr->path().string().c_str(), limg)) {
// take the left most square of the image
int h = limg.dim(0), w = limg.dim(1), newh, neww;
if (h > w) {
newh = width;
neww = (newh / (float) h) * w;
} else {
neww = width;
newh = (neww / (float) w) * h;
}
// resize image to target width
limg = image_resize(limg, neww, newh, 1);
left_images = images[current_ds];
left_images = left_images.narrow(3, channels_size, 0);
left_images = left_images.narrow(0, 1 + (deformations<=0?0:
deformations),
current_used);
current_labels = labels[current_ds];
current_labels = current_labels.narrow(0, current_labels.dim(0) -
current_used,current_used);
current_deformid = deformid[current_ds];
current_deformid = current_deformid.
narrow(0, current_deformid.dim(0) - current_used, current_used);
tmp = left_images[0];
// convert and copy image into images buffer
convert_image(limg, tmp, channels_mode, fkernel_size);
// if adding to dataset 0, add deformations if deformations > 0
if ((current_ds == 0) && (deformations > 0)) {
add_deformations(left_images, current_labels, label,
deformations, current_deformid, ndefid);
counters_used.set(counters_used.get(current_ds) + deformations,
current_ds);
} else // no deformations
left_images = left_images.narrow(0, 1, 0);
// display
if (display) display_images(classes, label, limg, left_images,
channels_mode, ds_names, current_ds);
// increment counter for dataset current_ds
counters_used.set(counters_used.get(current_ds) + 1, current_ds);
}
}
counter++;
}
}
}
}
void class_answer<T, Tds1, Tds2>::fprop1(idx<T> &in, idx<T> &out)
{
// resize out if necessary
idxdim d(in);
d.setdim(0, 2); // 2 outputs per pixel: class,confidence
idx<T> outx = out;
idx<T> inx = in;
if (resize_output)
{
if (d != out.get_idxdim())
{
out.resize(d);
outx = out;
}
}
else // if not resizing, narrow to the number of targets
{
if (outx.dim(0) != targets.dim(0))
outx = outx.narrow(0, targets.dim(0), 0);
}
// apply tanh if required
if (apply_tanh)
{
mtanh.fprop1(in, tmp);
inx = tmp;
}
// loop on features (dimension 0) to set class and confidence
int classid;
T conf, max2 = 0;
idx_1loop2(ii, inx, T, oo, outx, T, {
if (binary_target)
{
T t0 = targets.gget(0);
T t1 = targets.gget(1);
T a = ii.gget();
if (std::fabs((double)a - t0) < std::fabs((double)a - t1))
{
oo.set((T)0, 0); // class 0
oo.set((T)(2 - std::fabs((double)a - t0)) / 2, 1); // conf
}
else
{
oo.set((T)1, 0); // class 1
oo.set((T)(2 - std::fabs((double)a - t1)) / 2, 1); // conf
}
}
else if (single_output >= 0)
{
oo.set((T)single_output, 0); // all answers are the same class
oo.set((T)((ii.get(single_output) - target_min) / target_range), 1);
}
else // 1-of-n target
{ // set class answer
if (force_class >= 0) classid = force_class;
else classid = idx_indexmax(ii);
oo.set((T)classid, 0);
// set confidence
intg p;
bool ini = false;
switch (conf_type)
{
case confidence_sqrdist: // squared distance to target
target = targets.select(0, classid);
conf = (T)(1.0 - ((idx_sqrdist(target, ii) - conf_shift)
/ conf_ratio));
oo.set(conf, 1);
break;
case confidence_single: // simply return class' out (normalized)
conf = (T)((ii.get(classid) - conf_shift) / conf_ratio);
oo.set(conf, 1);
break;
case confidence_max: // distance with 2nd max answer
conf = std::max(target_min, std::min(target_max, ii.get(classid)));
for (p = 0; p < ii.dim(0); ++p)
{
if (p != classid)
{
if (!ini)
{
max2 = ii.get(p);
ini = true;
}
else
{
if (ii.get(p) > max2)
max2 = ii.get(p);
}
}
}
max2 = std::max(target_min, std::min(target_max, max2));
oo.set((T)(((conf - max2) - conf_shift) / conf_ratio), 1);
break;
default:
eblerror("confidence type " << conf_type << " undefined");
}
}
});
