void Net::Train(const mxArray *mx_data, const mxArray *mx_labels) {
//mexPrintMsg("Start training...");
ReadData(mx_data);
ReadLabels(mx_labels);
InitNorm();
std::srand(params_.seed_);
size_t train_num = labels_.size1();
size_t numbatches = (size_t) ceil((ftype) train_num/params_.batchsize_);
trainerror_.resize(params_.numepochs_, numbatches);
for (size_t epoch = 0; epoch < params_.numepochs_; ++epoch) {
std::vector<size_t> randind(train_num);
for (size_t i = 0; i < train_num; ++i) {
randind[i] = i;
}
if (params_.shuffle_) {
std::random_shuffle(randind.begin(), randind.end());
}
std::vector<size_t>::const_iterator iter = randind.begin();
for (size_t batch = 0; batch < numbatches; ++batch) {
size_t batchsize = std::min(params_.batchsize_, (size_t)(randind.end() - iter));
std::vector<size_t> batch_ind = std::vector<size_t>(iter, iter + batchsize);
iter = iter + batchsize;
Mat data_batch = SubMat(data_, batch_ind, 1);
Mat labels_batch = SubMat(labels_, batch_ind, 1);
UpdateWeights(epoch, false);
InitActiv(data_batch);
Mat pred_batch;
Forward(pred_batch, 1);
InitDeriv(labels_batch, trainerror_(epoch, batch));
Backward();
CalcWeights();
UpdateWeights(epoch, true);
if (params_.verbose_ == 2) {
std::string info = std::string("Epoch: ") + std::to_string(epoch+1) +
std::string(", batch: ") + std::to_string(batch+1);
mexPrintMsg(info);
}
} // batch
if (params_.verbose_ == 1) {
std::string info = std::string("Epoch: ") + std::to_string(epoch+1);
mexPrintMsg(info);
}
} // epoch
//mexPrintMsg("Training finished");
}
void Net::Train(const mxArray *mx_data, const mxArray *mx_labels) {
//mexPrintMsg("Start training...");
LayerFull *lastlayer = static_cast<LayerFull*>(layers_.back());
std::vector<size_t> labels_dim = mexGetDimensions(mx_labels);
mexAssert(labels_dim.size() == 2, "The label array must have 2 dimensions");
mexAssert(labels_dim[0] == lastlayer->length_,
"Labels and last layer must have equal number of classes");
size_t train_num = labels_dim[1];
Mat labels(labels_dim);
mexGetMatrix(mx_labels, labels);
classcoefs_.assign(labels_dim[0], 1);
if (params_.balance_) {
Mat labels_mean(labels_dim[0], 1);
labels.Mean(2, labels_mean);
for (size_t i = 0; i < labels_dim[0]; ++i) {
mexAssert(labels_mean(i) > 0, "Balancing impossible: one of the classes is not presented");
(classcoefs_[i] /= labels_mean(i)) /= labels_dim[0];
}
}
if (lastlayer->function_ == "SVM") {
(labels *= 2) -= 1;
}
size_t mapnum = 1;
if (mexIsCell(mx_data)) {
mapnum = mexGetNumel(mx_data);
}
mexAssert(mapnum == layers_.front()->outputmaps_,
"Data must have the same number of cells as outputmaps on the first layer");
std::vector< std::vector<Mat> > data(mapnum);
for (size_t map = 0; map < mapnum; ++map) {
const mxArray *mx_cell;
if (mexIsCell(mx_data)) {
mx_cell = mxGetCell(mx_data, map);
} else {
mx_cell = mx_data;
}
std::vector<size_t> data_dim = mexGetDimensions(mx_cell);
mexAssert(data_dim.size() == 3, "The data array must have 3 dimensions");
mexAssert(data_dim[0] == layers_.front()->mapsize_[0] &&
data_dim[1] == layers_.front()->mapsize_[1],
"Data and the first layer must have equal sizes");
mexAssert(data_dim[2] == train_num, "All data maps and labels must have equal number of objects");
mexGetMatrix3D(mx_cell, data[map]);
}
size_t numbatches = ceil((double) train_num/params_.batchsize_);
trainerror_.assign(params_.numepochs_ * numbatches, 0);
for (size_t epoch = 0; epoch < params_.numepochs_; ++epoch) {
std::vector<size_t> randind(train_num);
for (size_t i = 0; i < train_num; ++i) {
randind[i] = i;
}
if (params_.shuffle_) {
std::random_shuffle(randind.begin(), randind.end());
}
std::vector<size_t>::const_iterator iter = randind.begin();
for (size_t batch = 0; batch < numbatches; ++batch) {
size_t batchsize = std::min(params_.batchsize_, (size_t)(randind.end() - iter));
std::vector<size_t> batch_ind = std::vector<size_t>(iter, iter + batchsize);
iter = iter + batchsize;
std::vector< std::vector<Mat> > data_batch(mapnum);
for (size_t map = 0; map < mapnum; ++map) {
data_batch[map].resize(batchsize);
for (size_t i = 0; i < batchsize; ++i) {
data_batch[map][i] = data[map][batch_ind[i]];
}
}
Mat labels_batch(labels_dim[0], batchsize);
Mat pred_batch(labels_dim[0], batchsize);
labels.SubMat(batch_ind, 2 ,labels_batch);
UpdateWeights(false);
Forward(data_batch, pred_batch, true);
Backward(labels_batch, trainerror_[epoch * numbatches + batch]);
UpdateWeights(true);
if (params_.verbose_ == 2) {
std::string info = std::string("Epoch: ") + std::to_string(epoch+1) +
std::string(", batch: ") + std::to_string(batch+1);
mexPrintMsg(info);
}
} // batch
if (params_.verbose_ == 1) {
std::string info = std::string("Epoch: ") + std::to_string(epoch+1);
mexPrintMsg(info);
}
} // epoch
//mexPrintMsg("Training finished");
}
void mexFunction(int nLhs, mxArray* pLhs[], int nRhs, const mxArray* pRhs[]) {
mexAssert(NARGIN_MIN <= nRhs && nRhs <= NARGIN_MAX, "Number of input arguments in wrong!");
mexAssert(nLhs == NARGOUT, "Number of output arguments is wrong!" );
mexAssert(mexIsCell(IN_L), "Layers must be the cell array");
mexAssert(mexGetNumel(IN_L) == 2, "Layers array must contain 2 cells");
mexAssert(mexIsCell(IN_W), "Weights must be the cell array");
mexAssert(mexGetNumel(IN_W) == 2, "Weights array must contain 2 cells");
Net net;
mxArray *mx_weights;
net.InitLayers(mexGetCell(IN_L, 1));
net.InitWeights(mexGetCell(IN_W, 1), mx_weights);
net.InitParams(IN_P);
net.ReadLabels(IN_Y);
const mxArray *mx_imweights = mexGetCell(IN_W, 0);
size_t train_num = net.labels_.size1();
mexAssert(train_num == mexGetNumel(mx_imweights),
"Weights and labels number must coincide");
bool is_multicoords = false;
if (mexIsCell(IN_X)) {
mexAssert(train_num == mexGetNumel(IN_X),
"Coordinates and labels number must coincide");
is_multicoords = true;
}
Params params_ = net.params_;
size_t numbatches = (size_t) ceil((ftype) train_num/params_.batchsize_);
Mat trainerror_(params_.numepochs_, numbatches);
Mat trainerror2_(params_.numepochs_, numbatches);
trainerror2_.assign(0);
std::vector<Net> imnets;
imnets.resize(params_.batchsize_);
for (size_t i = 0; i < params_.batchsize_; ++i) {
imnets[i].InitLayers(mexGetCell(IN_L, 0));
if (!is_multicoords) {
imnets[i].ReadData(IN_X);
} else {
imnets[i].ReadData(mexGetCell(IN_X, i)); // just to get pixels_num
}
}
size_t pixels_num = imnets[0].data_.size1();
Layer *firstlayer = net.layers_[0];
size_t dimens_num = firstlayer->outputmaps_;
mexAssert(imnets[0].layers_.back()->length_ == dimens_num,
"Final layer length must coincide with the number of outputmaps");
mexAssert(pixels_num == firstlayer->mapsize_[0] * firstlayer->mapsize_[1],
"Pixels number must coincide with the first layer elements number");
std::vector<size_t> pred_size(2);
pred_size[0] = 1; pred_size[1] = pixels_num * dimens_num;
Mat images_mat, labels_batch, pred_batch, pred_pixels;
std::vector< std::vector<Mat> > images, images_der;
for (size_t epoch = 0; epoch < params_.numepochs_; ++epoch) {
std::vector<size_t> randind(train_num);
for (size_t i = 0; i < train_num; ++i) {
randind[i] = i;
}
if (params_.shuffle_) {
std::random_shuffle(randind.begin(), randind.end());
}
std::vector<size_t>::const_iterator iter = randind.begin();
for (size_t batch = 0; batch < numbatches; ++batch) {
size_t batchsize = std::min(params_.batchsize_, (size_t)(randind.end() - iter));
std::vector<size_t> batch_ind = std::vector<size_t>(iter, iter + batchsize);
iter = iter + batchsize;
labels_batch = SubMat(net.labels_, batch_ind, 1);
net.UpdateWeights(epoch, false);
images_mat.resize(batchsize, pred_size[1]);
InitMaps(images_mat, pred_size, images);
// first pass
for (size_t m = 0; m < batchsize; ++m) {
imnets[m].InitWeights(mexGetCell(mx_imweights, batch_ind[m]));
if (is_multicoords) {
imnets[m].ReadData(mexGetCell(IN_X, batch_ind[m]));
}
imnets[m].InitActiv(imnets[m].data_);
imnets[m].Forward(pred_pixels, 1);
images[m][0].copy(Trans(pred_pixels).reshape(pred_size[0], pred_size[1]));
}
net.InitActiv(images_mat);
net.Forward(pred_batch, 1);
/*
for (int i = 0; i < 5; ++i) {
mexPrintMsg("pred_batch1", pred_batch(0, i));
}*/
// second pass
net.InitDeriv(labels_batch, trainerror_(epoch, batch));
net.Backward();
net.CalcWeights();
InitMaps(firstlayer->deriv_mat_, pred_size, images_der);
for (size_t m = 0; m < batchsize; ++m) {
imnets[m].layers_.back()->deriv_mat_ = Trans(images_der[m][0].reshape(dimens_num, pixels_num));
imnets[m].Backward();
}
// third pass
ftype loss2 = 0, curloss = 0, invind = 0;
std::vector<size_t> invalid;
for (size_t m = 0; m < batchsize; ++m) {
imnets[m].InitDeriv2(curloss);
if (curloss > 0) {
imnets[m].Forward(pred_pixels, 3);
images[m][0].copy(Trans(pred_pixels).reshape(pred_size[0], pred_size[1]));
loss2 += curloss;
} else {
invalid.push_back(m);
}
}
if (invalid.size() < batchsize) {
loss2 /= (batchsize - invalid.size());
trainerror2_(epoch, batch) = loss2;
net.InitActiv(images_mat);
net.Forward(pred_batch, 3);
}
net.CalcWeights2(invalid);
// weights update
net.UpdateWeights(epoch, true);
if (params_.verbose_ == 2) {
std::string info = std::string("Epoch: ") + std::to_string(epoch+1) +
std::string(", batch: ") + std::to_string(batch+1);
mexPrintMsg(info);
}
} // batch
if (params_.verbose_ == 1) {
std::string info = std::string("Epoch: ") + std::to_string(epoch+1);
mexPrintMsg(info);
}
} // epoch
//mexPrintMsg("Training finished");
//net.weights_.get().copy(net.weights_.der());
OUT_W = mexSetCellMat(1, 2);
mexSetCell(OUT_W, 0, mexDuplicateArray(mx_imweights));
mexSetCell(OUT_W, 1, mx_weights);
OUT_E = mexSetCellMat(1, 2);
mexSetCell(OUT_E, 0, mexSetMatrix(trainerror_));
mexSetCell(OUT_E, 1, mexSetMatrix(trainerror2_));
}
void LayerJitt::Init(const mxArray *mx_layer, Layer *prev_layer) {
//mexAssert(prev_layer->type_ == "i", "The 'j' type layer must be after the input one");
numdim_ = prev_layer->numdim_;
outputmaps_ = prev_layer->outputmaps_;
length_prev_ = prev_layer->length_prev_;
mexAssert(mexIsField(mx_layer, "mapsize"), "The 'j' type layer must contain the 'mapsize' field");
std::vector<ftype> mapsize = mexGetVector(mexGetField(mx_layer, "mapsize"));
mexAssert(mapsize.size() == numdim_, "Length of jitter mapsize vector and maps dimensionality must coincide");
mapsize_.resize(numdim_);
length_ = outputmaps_;
for (size_t i = 0; i < numdim_; ++i) {
mexAssert(1 <= mapsize[i], "In 'j' layer mapsize must be positive");
mapsize_[i] = (size_t) mapsize[i];
length_ *= mapsize_[i];
}
shift_.assign(numdim_, 0);
if (mexIsField(mx_layer, "shift")) {
shift_ = mexGetVector(mexGetField(mx_layer, "shift"));
mexAssert(shift_.size() == numdim_, "Length of jitter shift vector and maps dimensionality must coincide");
for (size_t i = 0; i < numdim_; ++i) {
mexAssert(0 <= shift_[i] && shift_[i] < mapsize_[i], "Shift in 'j' layer is out of range");
}
}
scale_.assign(numdim_, 1);
if (mexIsField(mx_layer, "scale")) {
scale_ = mexGetVector(mexGetField(mx_layer, "scale"));
mexAssert(scale_.size() == numdim_, "Length of jitter scale vector and maps dimensionality must coincide");
for (size_t i = 0; i < numdim_; ++i) {
mexAssert(1 <= scale_[i] && scale_[i] < mapsize_[i], "Scale in 'j' layer is out of range");
}
}
mirror_.assign(numdim_, false);
if (mexIsField(mx_layer, "mirror")) {
std::vector<ftype> mirror = mexGetVector(mexGetField(mx_layer, "mirror"));
mexAssert(mirror.size() == numdim_, "Length of jitter scale vector and maps dimensionality must coincide");
for (size_t i = 0; i < numdim_; ++i) {
mirror_[i] = (mirror[i] > 0);
}
}
angle_ = 0;
if (mexIsField(mx_layer, "angle")) {
angle_ = mexGetScalar(mexGetField(mx_layer, "angle"));
mexAssert(0 <= angle_ && angle_ <= 1, "Angle in 'j' layer must be between 0 and 1");
}
defval_ = 0;
if (mexIsField(mx_layer, "defval")) {
defval_ = mexGetScalar(mexGetField(mx_layer, "defval"));
} else {
// check that the transformed image is always inside the original one
std::vector<ftype> maxsize(numdim_, 0);
for (size_t i = 0; i < numdim_; ++i) {
maxsize[i] = (ftype) (mapsize_[i] - 1) * scale_[i];
}
if (angle_ > 0) {
ftype angle_inn = atan2((ftype) mapsize_[0], (ftype) mapsize_[1]) / kPi;
ftype maxsin = 1;
if (angle_inn + angle_ < 0.5) {
maxsin = sin(kPi * (angle_inn + angle_));
}
ftype maxcos = 1;
if (angle_inn > angle_) {
maxcos = cos(kPi * (angle_inn - angle_));
}
ftype maxrad = sqrt(maxsize[0]*maxsize[0] + maxsize[1]*maxsize[1]);
maxsize[0] = maxrad * maxsin;
maxsize[1] = maxrad * maxcos;
}
std::vector<ftype> oldmapsize(numdim_, 0);
for (size_t i = 0; i < numdim_; ++i) {
oldmapsize[i] = (ftype) prev_layer->mapsize_[i];
}
ftype min0 = (oldmapsize[0]/2 - 0.5) - maxsize[0]/2 - shift_[0];
ftype max0 = (oldmapsize[0]/2 - 0.5) + maxsize[0]/2 + shift_[0];
ftype min1 = (oldmapsize[1]/2 - 0.5) - maxsize[1]/2 - shift_[1];
ftype max1 = (oldmapsize[1]/2 - 0.5) + maxsize[1]/2 + shift_[1];
if (!(0 <= min0 && max0 < oldmapsize[0] && 0 <= min1 && max1 < oldmapsize[1])) {
mexPrintMsg("min1", min0); mexPrintMsg("max1", max0);
mexPrintMsg("min2", min1); mexPrintMsg("max2", max1);
mexAssert(false, "For these jitter parameters the new image is out of the original image");
}
}
}
void LayerJitt::Init(const mxArray *mx_layer, const Layer *prev_layer) {
dims_[1] = prev_layer->dims_[1];
std::vector<ftype> shift(2);
shift[0] = 0; shift[1] = 0;
if (mexIsField(mx_layer, "shift")) {
shift = mexGetVector(mexGetField(mx_layer, "shift"));
mexAssertMsg(shift.size() == 2, "Length of jitter shift vector and maps dimensionality must coincide");
for (size_t i = 0; i < 2; ++i) {
mexAssertMsg(0 <= shift[i] && shift[i] < dims_[i+2], "Shift in 'jitt' layer is out of range");
}
MatCPU shift_cpu(1, 2);
shift_cpu.assign(shift);
shift_ = shift_cpu;
}
std::vector<ftype> scale(2);
scale[0] = 1; scale[1] = 1;
if (mexIsField(mx_layer, "scale")) {
scale = mexGetVector(mexGetField(mx_layer, "scale"));
mexAssertMsg(scale.size() == 2, "Length of jitter scale vector and maps dimensionality must coincide");
for (size_t i = 0; i < 2; ++i) {
mexAssertMsg(1 <= scale[i] && scale[i] < dims_[i+2], "Scale in 'j' layer is out of range");
}
MatCPU scale_cpu(1, 2);
scale_cpu.assign(scale);
scale_ = scale_cpu;
scale_.Log();
}
if (mexIsField(mx_layer, "mirror")) {
std::vector<ftype> mirror = mexGetVector(mexGetField(mx_layer, "mirror"));
mexAssertMsg(mirror.size() == 2, "Length of jitter scale vector and maps dimensionality must coincide");
for (size_t i = 0; i < 2; ++i) {
mexAssertMsg(mirror[i] == 0 || mirror[i] == 1, "Mirror must be either 0 or 1");
}
MatCPU mirror_cpu(1, 2);
mirror_cpu.assign(mirror);
mirror_ = mirror_cpu;
}
if (mexIsField(mx_layer, "angle")) {
angle_ = mexGetScalar(mexGetField(mx_layer, "angle"));
mexAssertMsg(0 <= angle_ && angle_ <= 1, "Angle in 'j' layer must be between 0 and 1");
}
if (mexIsField(mx_layer, "defval")) {
defval_ = mexGetScalar(mexGetField(mx_layer, "defval"));
} else {
// check that the transformed image is always inside the original one
std::vector<ftype> maxsize(2, 0);
for (size_t i = 0; i < 2; ++i) {
maxsize[i] = (ftype) (dims_[i+2] - 1) * scale[i];
}
if (angle_ > 0) {
ftype angle_inn = atan2((ftype) dims_[2], (ftype) dims_[3]) / kPi;
ftype maxsin = 1;
if (angle_inn + angle_ < 0.5) {
maxsin = sin(kPi * (angle_inn + angle_));
}
ftype maxcos = 1;
if (angle_inn > angle_) {
maxcos = cos(kPi * (angle_inn - angle_));
}
ftype maxrad = (ftype) sqrt((double) (maxsize[0]*maxsize[0] + maxsize[1]*maxsize[1]));
maxsize[0] = maxrad * maxsin;
maxsize[1] = maxrad * maxcos;
}
std::vector<ftype> oldmapsize(2, 0);
for (size_t i = 0; i < 2; ++i) {
oldmapsize[i] = (ftype) prev_layer->dims_[i+2];
}
ftype min0 = ((ftype) oldmapsize[0] / 2 - (ftype) 0.5) - (ftype) maxsize[0] / 2 - shift[0];
ftype max0 = ((ftype) oldmapsize[0] / 2 - (ftype) 0.5) + (ftype) maxsize[0] / 2 + shift[0];
ftype min1 = ((ftype) oldmapsize[1] / 2 - (ftype) 0.5) - (ftype) maxsize[1] / 2 - shift[1];
ftype max1 = ((ftype) oldmapsize[1] / 2 - (ftype) 0.5) + (ftype) maxsize[1] / 2 + shift[1];
if (!(0 <= min0 && max0 < oldmapsize[0] && 0 <= min1 && max1 < oldmapsize[1])) {
mexPrintMsg("min1", min0); mexPrintMsg("max1", max0);
mexPrintMsg("min2", min1); mexPrintMsg("max2", max1);
mexAssertMsg(false, "For these jitter parameters the new image is out of the original image");
}
}
if (mexIsField(mx_layer, "eigenvectors")) {
const mxArray* mx_ev = mexGetField(mx_layer, "eigenvectors");
std::vector<size_t> ev_dim = mexGetDimensions(mx_ev);
mexAssertMsg(ev_dim.size() == 2, "The eigenvectors array must have 2 dimensions");
mexAssertMsg(ev_dim[0] == dims_[1] && ev_dim[1] == dims_[1],
"The eigenvector matrix size is wrong");
MatCPU ev_cpu(dims_[1], dims_[1]);
mexGetMatrix(mx_ev, ev_cpu);
eigenvectors_.resize(dims_[1], dims_[1]);
eigenvectors_ = ev_cpu;
if (mexIsField(mx_layer, "noise_std")) {
noise_std_ = mexGetScalar(mexGetField(mx_layer, "noise_std"));
mexAssertMsg(noise_std_ >= 0, "noise_std must be nonnegative");
} else {
mexAssertMsg(false, "noise_std is required with eigenvalues");
}
}
if (mexIsField(mx_layer, "randtest")) {
randtest_ = (mexGetScalar(mexGetField(mx_layer, "randtest")) > 0);
}
}
