Shark Machine Learning Library
FFNet.cpp
Go to the documentation of this file.00001 //=========================================================================== 00042 //=========================================================================== 00043 00044 #include <SharkDefs.h> 00045 #include <ReClaM/FFNet.h> 00046 #include <sstream> 00047 00048 00049 using namespace std; 00050 00051 //=========================================================================== 00067 FFNet::~FFNet() 00068 { 00069 if (z) delete [] z; 00070 } 00071 00072 //=========================================================================== 00111 double FFNet::g(double a) 00112 { 00113 return 1 / (1 + exp(-a)); 00114 } 00115 00116 00117 //=========================================================================== 00144 double FFNet::dg(double ga) 00145 { 00146 return ga *(1 - ga); 00147 } 00148 00149 00150 //=========================================================================== 00190 double FFNet::gOutput(double a) 00191 { 00192 return 1 / (1 + exp(-a)); 00193 } 00194 00195 //=========================================================================== 00223 double FFNet::dgOutput(double ga) 00224 { 00225 return ga *(1 - ga); 00226 } 00227 00228 //=========================================================================== 00254 void FFNet::initWeights(long seed, double l, double h) 00255 { 00256 Rng::seed(seed); 00257 for (unsigned i = inputDimension; i < numberOfNeurons; i++) 00258 { 00259 for (unsigned j = 0; j < i; j++) 00260 if (connection(i, j)) weight(i, j) = Rng::uni(l, h); 00261 if (connectionMatrix(i, bias)) weight(i, bias) = Rng::uni(l, h); 00262 } 00263 writeParameters(); 00264 } 00265 00266 //=========================================================================== 00291 inline 00292 void FFNet::initWeights(double l, double h) 00293 { 00294 for (unsigned i = inputDimension; i < numberOfNeurons; i++) 00295 { 00296 for (unsigned j = 0; j < i; j++) 00297 if (connection(i, j)) weight(i, j) = Rng::uni(l, h); 00298 if (connectionMatrix(i, bias)) weight(i, bias) = Rng::uni(l, h); 00299 } 00300 writeParameters(); 00301 } 00302 00303 //=========================================================================== 00329 inline 00330 void FFNet::model(const Array<double> &input, Array<double> &output) 00331 { 00332 readParameters(); 00333 if (input.ndim() == 1) 00334 { 00335 output.resize(outputDimension, false); 00336 activate(input); 00337 for (unsigned i = firstOutputNeuron, c = 0 ; i < numberOfNeurons; i++, c++) 00338 output(c) = z[i]; 00339 } 00340 else 00341 { 00342 output.resize(input.dim(0), outputDimension, false); 00343 for (unsigned pattern = 0; pattern < input.dim(0); pattern++) 00344 { 00345 activate(input[pattern]); 00346 for (unsigned i = firstOutputNeuron, c = 0 ; i < numberOfNeurons; i++, c++) 00347 output(pattern, c) = z[i]; 00348 } 00349 } 00350 } 00351 00352 //=========================================================================== 00374 inline 00375 void FFNet::backprop(Array<double>& derivative) 00376 { 00377 unsigned i, j; // denote neurons 00378 unsigned c; // denotes output 00379 unsigned pos; // the actual position in the Array derivative 00380 double sum; 00381 00382 // 1. the 'local gradient' deltas for each 00383 // output are calculated 00384 00385 // output neurons 00386 for (j = firstOutputNeuron; j < numberOfNeurons; j++) 00387 for (c = 0; c < outputDimension; c++) 00388 if (c + firstOutputNeuron == j) d(c, j) = dgOutput(z[j]); 00389 else d(c, j) = 0; 00390 00391 // hidden units 00392 for (j = firstOutputNeuron - 1; j >= inputDimension; j--) 00393 { 00394 for (c = 0; c < outputDimension; c++) 00395 { 00396 sum = 0; 00397 for (i = j + 1; i < numberOfNeurons; i++) 00398 { 00399 if (connection(i, j)) sum += weight(i, j) * d(c, i); 00400 } 00401 d(c, j) = dg(z[j]) * sum; 00402 } 00403 } 00404 00405 // 2. the gradient for each output is calculated from 00406 // delta 00407 pos = 0; 00408 for (i = inputDimension; i < numberOfNeurons; i++) 00409 { 00410 for (j = 0; j < i; j++) 00411 { 00412 if (connection(i, j)) 00413 { 00414 for (c = 0; c < outputDimension; c++) 00415 { 00416 derivative(c, pos) += d(c, i) * z[j]; 00417 } 00418 pos++; 00419 } 00420 } 00421 if (connection(i, bias)) 00422 { 00423 for (c = 0; c < outputDimension; c++) 00424 { // bias 00425 derivative(c, pos) += d(c, i); 00426 } 00427 pos++; 00428 } 00429 } 00430 } 00431 00432 //=========================================================================== 00457 void FFNet::modelDerivative(const Array<double> &input, Array<double> &output, Array<double>& derivative) 00458 { 00459 unsigned i; // denotes neurons 00460 unsigned c; // denotes output 00461 00462 derivative.resize(getOutputDimension(), getParameterDimension(), false); 00463 derivative = 0.; 00464 00465 readParameters(); 00466 00467 if (input.ndim() == 1) 00468 { 00469 output.resize(getOutputDimension(), false); 00470 00471 // calculate output 00472 activate(input); 00473 00474 for (i = firstOutputNeuron, c = 0 ; i < numberOfNeurons; i++, c++) { 00475 output(c) = z[i]; 00476 } 00477 00478 // backpropagation step 00479 backprop(derivative); 00480 } 00481 else 00482 { 00483 output.resize(input.dim(0), getOutputDimension(), false); 00484 for (unsigned pattern = 0; pattern < input.dim(0); pattern++) 00485 { 00486 activate(input[pattern]); 00487 for (i = firstOutputNeuron, c = 0 ; i < numberOfNeurons; i++, c++) { 00488 output(pattern, c) = z[i]; 00489 } 00490 backprop(derivative); 00491 } 00492 } 00493 }; 00494 00495 void FFNet::generalDerivative(const Array<double>& input, const Array<double>& coeffs, Array<double>& derivative) 00496 { 00497 unsigned i, j; // denote neurons 00498 unsigned c; // denotes output 00499 unsigned pos; // the actual position in the Array dedw 00500 double sum; 00501 00502 for (i = firstOutputNeuron, c = 0 ; i < numberOfNeurons; i++, c++) 00503 { 00504 delta(i) = coeffs(c) * dgOutput(z[i]); 00505 } 00506 00507 // hidden units 00508 for (j = firstOutputNeuron - 1; j >= inputDimension; j--) 00509 { 00510 sum = 0; 00511 for (i = j + 1; i < numberOfNeurons; i++) 00512 { 00513 if (connection(i, j)) sum += weight(i, j) * delta(i); 00514 } 00515 delta(j) = dg(z[j]) * sum; 00516 } 00517 00518 // calculate error gradient 00519 pos = 0; 00520 for (i = inputDimension; i < numberOfNeurons; i++) 00521 { 00522 for (j = 0; j < i; j++) 00523 { 00524 if (connection(i, j)) 00525 { 00526 derivative(pos) -= delta(i) * z[j]; 00527 pos++; 00528 } 00529 } 00530 if (connection(i, bias)) 00531 { // bias 00532 derivative(pos) -= delta(i); 00533 pos++; 00534 } 00535 } 00536 }; 00537 00538 //=========================================================================== 00566 void FFNet::modelDerivative(const Array<double>& input, Array<double>& derivative) 00567 { 00568 readParameters(); 00569 00570 derivative.resize(getOutputDimension(), getParameterDimension(), false); 00571 derivative = 0.; 00572 00573 if (input.ndim() == 1) 00574 { 00575 activate(input); 00576 // backpropagation step 00577 backprop(derivative); 00578 } 00579 else 00580 { 00581 stringstream s; 00582 s << "the derivative of the model with respect to more than one input is not defined" << endl; 00583 throw SHARKEXCEPTION(s.str().c_str()); 00584 } 00585 }; 00586 00587 00588 //=========================================================================== 00612 void FFNet::resize() 00613 { 00614 numberOfNeurons = connectionMatrix.dim(0); 00615 bias = numberOfNeurons; 00616 firstOutputNeuron = numberOfNeurons - outputDimension; 00617 00618 unsigned numberOfWeights = 0; //numberOfNeurons - inputDimension; 00619 for (unsigned i = 0; i < connectionMatrix.dim(0); i++) 00620 { 00621 for (unsigned j = 0; j < i; j++) 00622 if (connectionMatrix(i, j)) numberOfWeights++; 00623 if (connectionMatrix(i, bias)) numberOfWeights++; 00624 } 00625 00626 parameter.resize(numberOfWeights, false); 00627 d.resize(outputDimension, numberOfNeurons, false); 00628 delta.resize(numberOfNeurons, false); 00629 00630 if (z) delete [] z; 00631 if (numberOfNeurons) z = new double[numberOfNeurons]; 00632 else z = NULL; 00633 } 00634 00635 00636 //=========================================================================== 00660 FFNet::FFNet(const unsigned in, const unsigned out) 00661 { 00662 inputDimension = in; 00663 outputDimension = out; 00664 00665 numberOfNeurons = 0; 00666 bias = 0; 00667 firstOutputNeuron = 0; 00668 00669 z = NULL; 00670 } 00671 00672 00673 //=========================================================================== 00702 FFNet::FFNet(const unsigned in, unsigned out, 00703 const Array<int>& cmat) 00704 { 00705 if (cmat.ndim() != 2) 00706 { 00707 stringstream s; 00708 s << "connection matrix has not dimension two" << endl; 00709 throw SHARKEXCEPTION(s.str().c_str()); 00710 } 00711 if (cmat.dim(0) != (cmat.dim(1) - 1)) 00712 { 00713 stringstream s; 00714 s << "connection matrix has to be (N + 1) x N" << endl; 00715 throw SHARKEXCEPTION(s.str().c_str()); 00716 } 00717 00718 inputDimension = in; 00719 outputDimension = out; 00720 connectionMatrix = cmat; 00721 00722 weightMatrix.resize(connectionMatrix, false); 00723 weightMatrix = 0; 00724 00725 z = NULL; 00726 00727 resize(); 00728 writeParameters(); 00729 } 00730 00731 00732 //=========================================================================== 00760 FFNet::FFNet(const unsigned in, unsigned out, 00761 const Array<int>& cmat, const Array<double>& wmat) 00762 { 00763 if (cmat.ndim() != 2) 00764 { 00765 stringstream s; 00766 s << "connection matrix has not dimension two" << endl; 00767 throw SHARKEXCEPTION(s.str().c_str()); 00768 } 00769 if (cmat.dim(0) != (cmat.dim(1) - 1)) 00770 { 00771 stringstream s; 00772 s << "connection matrix has to be (N + 1) x N" << endl; 00773 throw SHARKEXCEPTION(s.str().c_str()); 00774 } 00775 if (wmat.ndim() != 2) 00776 { 00777 stringstream s; 00778 s << "weight matrix has not dimension two" << endl; 00779 throw SHARKEXCEPTION(s.str().c_str()); 00780 } 00781 if (wmat.dim(0) != (wmat.dim(1) - 1)) 00782 { 00783 stringstream s; 00784 s << "weight matrix has to be (N + 1) x N" << endl; 00785 throw SHARKEXCEPTION(s.str().c_str()); 00786 } 00787 00788 inputDimension = in; 00789 outputDimension = out; 00790 connectionMatrix = cmat; 00791 weightMatrix = wmat; 00792 00793 z = NULL; 00794 resize(); 00795 writeParameters(); 00796 } 00797 00798 00799 //=========================================================================== 00851 FFNet::FFNet(const string &filename) 00852 { 00853 ifstream input(filename.c_str()); 00854 if (!input) 00855 { 00856 stringstream s; 00857 s << "cannot open net file " << filename << endl; 00858 throw SHARKEXCEPTION(s.str().c_str()); 00859 } 00860 z = NULL; 00861 input >> *this; 00862 input.close(); 00863 } 00864 00865 00866 //=========================================================================== 00893 void FFNet::setStructure(const Array<int>& cmat, const Array<double>& wmat) 00894 { 00895 if ((cmat.ndim() != 2) || 00896 (wmat.ndim() != 2)) 00897 { 00898 stringstream s; 00899 s << "connection or weight matrix has not dimension two" << endl; 00900 throw SHARKEXCEPTION(s.str().c_str()); 00901 } 00902 00903 if ((cmat.dim(0) != (cmat.dim(1) - 1)) || 00904 (cmat.dim(0) != wmat.dim(0)) || 00905 (cmat.dim(1) != wmat.dim(1))) 00906 { 00907 stringstream s; 00908 s << "connection and weight matrices have to be (N + 1) x N" << endl; 00909 throw SHARKEXCEPTION(s.str().c_str()); 00910 } 00911 00912 connectionMatrix = cmat; 00913 weightMatrix = wmat; 00914 00915 resize(); 00916 writeParameters(); 00917 } 00918 00919 00920 00921 //=========================================================================== 00957 void FFNet::setStructure(Array<unsigned> &layers, 00958 bool ff_layer, // all connections between layers? 00959 bool ff_in_out, // shortcuts from in to out? 00960 bool ff_all, // all shortcuts? 00961 bool bias) // bias? 00962 { 00963 unsigned N = 0; // total number of neurons: 00964 unsigned row, // connection matrix row (target neuron) 00965 column, // connection matrix column (start neuron) 00966 k, // counter. 00967 z_pos, // first target neuron in next layer 00968 s_pos; // first start neuron in current layer 00969 00970 00971 // 00972 // Calculate total number of neurons from the 00973 // number of neurons per layer: 00974 // 00975 for (k = 0; k < layers.dim(0); k++) N += layers(k); 00976 connectionMatrix.resize(N, N + 1); 00977 connectionMatrix = 0; 00978 00979 // 00980 // set connections from each neuron of layer i to each 00981 // neuron of layer i + 1 for all layers: 00982 // 00983 if (ff_layer) 00984 { 00985 z_pos = layers(0); 00986 s_pos = 0; 00987 for (k = 0; k < layers.dim(0) - 1; k++) 00988 { 00989 for (row = z_pos; row < z_pos + layers(k + 1); row++) 00990 for (column = s_pos; column < s_pos + layers(k); column++) 00991 connectionMatrix(row, column) = 1; 00992 s_pos += layers(k); 00993 z_pos += layers(k + 1); 00994 } 00995 } 00996 00997 // 00998 // set connections from all input neurons to all output neurons: 00999 // 01000 if (ff_in_out) 01001 { 01002 for (row = N - layers(layers.dim(0) - 1); row < N; row++) 01003 for (column = 0; column < layers(0); column++) 01004 connectionMatrix(row, column) = 1; 01005 } 01006 01007 // 01008 // set connections from all neurons of layer i to 01009 // all neurons of layers j with j > i for all layers i: 01010 // 01011 if (ff_all) 01012 { 01013 z_pos = layers(0); 01014 s_pos = 0; 01015 for (k = 0; k < layers.dim(0) - 1; k++) 01016 { 01017 for (row = z_pos; row < z_pos + layers(k + 1); row++) 01018 for (column = 0; column < s_pos + layers(k); column++) 01019 connectionMatrix(row, column) = 1; 01020 s_pos += layers(k); 01021 z_pos += layers(k + 1); 01022 } 01023 } 01024 01025 // 01026 // set connections from all neurons (except the input neurons) 01027 // to the bias values: 01028 // 01029 if (bias) 01030 for (k = layers(0); k < N; k++) connectionMatrix(k, N) = 1; 01031 01032 weightMatrix.resize(connectionMatrix.dim(0), connectionMatrix.dim(1), false); 01033 weightMatrix = 0.; 01034 01035 01036 inputDimension = layers(0); 01037 outputDimension = layers(layers.dim(0) - 1); 01038 01039 resize(); 01040 parameter = 0; 01041 01042 writeParameters(); 01043 } 01044 01045 01046 //=========================================================================== 01086 void FFNet::setStructure(unsigned in, unsigned hidden, unsigned out, 01087 bool ff_layer, // all connections beteween layers? 01088 bool ff_in_out, // shortcuts from in to out? 01089 bool ff_all, // all shortcuts? 01090 bool bias) // bias? 01091 { 01092 Array<unsigned> layer(3); 01093 layer(0) = in; 01094 layer(1) = hidden; 01095 layer(2) = out; 01096 setStructure(layer, ff_layer, ff_in_out, ff_all, bias); 01097 } 01098 01099 01100 //=========================================================================== 01141 void FFNet::setStructure(unsigned in, unsigned hidden1, unsigned hidden2, unsigned out, 01142 bool ff_layer, // all connections beteween layers? 01143 bool ff_in_out, // shortcuts from in to out? 01144 bool ff_all, // all shortcuts? 01145 bool bias) // bias? 01146 { 01147 Array<unsigned> layer(4); 01148 layer(0) = in; 01149 layer(1) = hidden1; 01150 layer(2) = hidden2; 01151 layer(3) = out; 01152 setStructure(layer, ff_layer, ff_in_out, ff_all, bias); 01153 } 01154 01155 //=========================================================================== 01180 void FFNet::setStructure(const Array<int>& cmat) 01181 { 01182 if (cmat.ndim() != 2) 01183 { 01184 stringstream s; 01185 s << "connection matrix has not dimension two" << endl; 01186 throw SHARKEXCEPTION(s.str().c_str()); 01187 } 01188 if (cmat.dim(0) != (cmat.dim(1) - 1)) 01189 { 01190 stringstream s; 01191 s << "connection matrix has to be (N + 1) x N" << endl; 01192 throw SHARKEXCEPTION(s.str().c_str()); 01193 } 01194 01195 connectionMatrix = cmat; 01196 weightMatrix.resize(cmat.dim(0), cmat.dim(1), false); 01197 01198 resize(); 01199 parameter = 0; 01200 01201 writeParameters(); 01202 } 01203 01204 01205 //=========================================================================== 01239 void FFNet::setStructure(const Array<double>& wmat, bool preserve) 01240 { 01241 if (wmat.ndim() != 2) 01242 { 01243 stringstream s; 01244 s << "weight matrix has not dimension two" << endl; 01245 throw SHARKEXCEPTION(s.str().c_str()); 01246 } 01247 if (wmat.dim(0) != (wmat.dim(1) - 1)) 01248 { 01249 stringstream s; 01250 s << "weight matrix has to be (N + 1) x N" << endl; 01251 throw SHARKEXCEPTION(s.str().c_str()); 01252 } 01253 01254 weightMatrix = wmat; 01255 connectionMatrix.resize(wmat.dim(0), wmat.dim(1), false); 01256 01257 numberOfNeurons = connectionMatrix.dim(0); 01258 bias = numberOfNeurons; 01259 firstOutputNeuron = numberOfNeurons - outputDimension; 01260 01261 unsigned numberOfWeights = 0; 01262 for (unsigned i = 0; i < connectionMatrix.dim(0); i++) 01263 { 01264 for (unsigned j = 0; j < i; j++) 01265 { 01266 if (preserve) 01267 { 01268 if (weightMatrix(i, j) != 0.) 01269 { 01270 if (connectionMatrix(i, j) == 0) 01271 { 01272 stringstream s; 01273 s << "weight specified for non-existing connection" << endl; 01274 throw SHARKEXCEPTION(s.str().c_str()); 01275 } 01276 numberOfWeights++; 01277 } 01278 else if (connectionMatrix(i, j) != 0) numberOfWeights++; 01279 if (weightMatrix(i, bias) != 0.) 01280 { 01281 if (connectionMatrix(i, bias) == 0) 01282 { 01283 stringstream s; 01284 s << "weight specified for non-existing bias" << endl; 01285 throw SHARKEXCEPTION(s.str().c_str()); 01286 } 01287 numberOfWeights++; 01288 } 01289 else if (connectionMatrix(i, bias) != 0) numberOfWeights++; 01290 } 01291 else 01292 { 01293 if (weightMatrix(i, j) != 0.) 01294 { 01295 connectionMatrix(i, j) = 1; 01296 numberOfWeights++; 01297 } 01298 else 01299 { 01300 connectionMatrix(i, j) = 0; 01301 } 01302 if (weightMatrix(i, bias) != 0.) 01303 { 01304 connectionMatrix(i, bias) = 1; 01305 numberOfWeights++; 01306 } 01307 else 01308 { 01309 connectionMatrix(i, bias) = 0; 01310 } 01311 } 01312 } 01313 } 01314 01315 resize(); 01316 writeParameters(); 01317 } 01318 01319 //=========================================================================== 01346 void FFNet::writeParameters() 01347 { 01348 for (unsigned k = 0, i = 0; i < connectionMatrix.dim(0); i++) 01349 { 01350 for (unsigned j = 0; j < i; j++) 01351 { 01352 if (connectionMatrix(i, j)) 01353 { 01354 parameter(k) = weight(i, j); 01355 k++; 01356 } 01357 } 01358 if (i >= inputDimension) 01359 { 01360 if (connectionMatrix(i, bias)) 01361 { 01362 parameter(k) = weight(i, bias); 01363 k++; 01364 } 01365 } 01366 } 01367 } 01368 01369 //=========================================================================== 01395 void FFNet::readParameters() 01396 { 01397 for (unsigned k = 0, i = 0; i < connectionMatrix.dim(0); i++) 01398 { 01399 for (unsigned j = 0; j < i; j++) 01400 { 01401 if (connectionMatrix(i, j)) 01402 { 01403 weight(i, j) = parameter(k); 01404 k++; 01405 } 01406 } 01407 if (i >= inputDimension) 01408 { 01409 if (connectionMatrix(i, bias)) 01410 { 01411 weight(i, bias) = parameter(k); 01412 k++; 01413 } 01414 } 01415 } 01416 } 01417 01418 01419 //=========================================================================== 01439 const Array< double > &FFNet::getWeights() 01440 { 01441 readParameters(); 01442 return weightMatrix; 01443 } 01444 01445 //=========================================================================== 01461 const Array< int > &FFNet::getConnections() 01462 { 01463 return connectionMatrix; 01464 } 01465 01466 //=========================================================================== 01490 void FFNet::write(ostream& os) const 01491 { 01492 os << inputDimension << " " << outputDimension << "\n"; 01493 writeArray(connectionMatrix, os, "", "\n", ' '); 01494 writeArray(weightMatrix, os, "", "\n", ' '); 01495 } 01496 01497 01498 //=========================================================================== 01525 void FFNet::read(istream& is) 01526 { 01527 unsigned i, j, n, pos; 01528 double v, dn; 01529 vector<double> l; 01530 01531 is >> inputDimension; 01532 is >> outputDimension; 01533 01534 is >> v; 01535 while (is) 01536 { 01537 l.push_back(v); 01538 is >> v; 01539 } 01540 01541 if (!l.size()) 01542 { 01543 stringstream s; 01544 s << "no matrices given in net file"; 01545 throw SHARKEXCEPTION(s.str().c_str()); 01546 } 01547 01548 dn = (sqrt(1. + l.size() * 2.) - 1.) / 2.; 01549 n = unsigned(dn); 01550 if (double(n) == dn) 01551 { // new style with bias 01552 connectionMatrix.resize(n, n + 1, false); 01553 weightMatrix.resize(n, n + 1, false); 01554 01555 pos = 0; 01556 for (i = 0; i < n; i++) 01557 { 01558 for (j = 0; j < n + 1; j++) 01559 { 01560 connectionMatrix(i, j) = unsigned(l[pos]); 01561 pos++; 01562 } 01563 } 01564 01565 for (i = 0; i < n; i++) 01566 { 01567 for (j = 0; j < n + 1; j++) 01568 { 01569 weightMatrix(i, j) = l[pos]; 01570 pos++; 01571 } 01572 } 01573 } 01574 else 01575 { // old style without bias 01576 dn = (sqrt(1. + l.size() * 8.) - 1.) / 4.; 01577 n = unsigned(dn); 01578 if (double(n) == dn) 01579 { 01580 connectionMatrix.resize(n, n + 1, false); 01581 weightMatrix.resize(n, n + 1, false); 01582 01583 pos = 0; 01584 for (i = 0; i < n; i++) 01585 { 01586 for (j = 0; j < n; j++) 01587 { 01588 connectionMatrix(i, j) = unsigned(l[pos]); 01589 pos++; 01590 } 01591 // handle bias 01592 if (i >= inputDimension) connectionMatrix(i, n) = 1; 01593 else connectionMatrix(i, n) = 0; 01594 } 01595 for (i = 0; i < n; i++) 01596 { 01597 for (j = 0; j < n + 1; j++) 01598 { 01599 weightMatrix(i, j) = l[pos]; 01600 pos++; 01601 } 01602 } 01603 } 01604 else 01605 { // no style 01606 stringstream s; 01607 s << "cannot parse network configuration file" << endl; 01608 throw SHARKEXCEPTION(s.str().c_str()); 01609 } 01610 } 01611 01612 resize(); 01613 writeParameters(); 01614 } 01615 01616 01617 void FFNet::replaceInString(const std::string &str, std::string &newString, const std::string &token, const std::string &newToken1, int newToken2, const std::string &newToken3) 01618 { 01619 std::ostringstream buffer; 01620 01621 buffer << newToken2; 01622 std::string newToken = newToken1 + buffer.str() + newToken3; 01623 01624 std::string::size_type pos; 01625 newString = str; 01626 01627 pos = newString.find(token, 0); 01628 while (pos != std::string::npos) 01629 { 01630 newString.replace(pos, token.length(), newToken); 01631 pos = newString.find(token, 0); 01632 } 01633 } 01634 01635 void FFNet::writeSource(std::ostream &os, const char *g, const char *gOut, unsigned p) 01636 { 01637 unsigned i, j; 01638 bool first; 01639 std::string str; 01640 01641 os.precision(p); 01642 01643 os << "void model(double *in, double *out) {\n"; 01644 for (j = 0; j < inputDimension; j++) 01645 os << " double z" << j << " = in[" << j << "];\n"; 01646 01647 for (i = inputDimension; i < firstOutputNeuron; i++) 01648 { 01649 first = true; 01650 for (j = 0; j < i; j++) 01651 if (connectionMatrix(i, j)) 01652 { 01653 if (first) 01654 { 01655 os << " double z" << i << " = z" << j << " * " << weightMatrix(i, j); 01656 first = false; 01657 } 01658 else 01659 { 01660 os << " + z" << j << " * " << weightMatrix(i, j); 01661 } 01662 } 01663 01664 if( connectionMatrix(i, bias) ) { 01665 if (first) 01666 { 01667 os << "double z" << i << " = " << weightMatrix(i, bias); // bias 01668 first = false; 01669 } 01670 else 01671 { 01672 os << " + " << weightMatrix(i, bias); 01673 } 01674 } 01675 01676 if (!first) 01677 { 01678 os << ";\n"; 01679 replaceInString(g, str, "#", "z", i, ""); 01680 os << " z" << i << " = " << str << ";\n"; 01681 } 01682 } 01683 for (i = firstOutputNeuron; i < numberOfNeurons; i++) 01684 { 01685 first = true; 01686 01687 for (j = 0; j < i; j++) 01688 if (connectionMatrix(i, j)) 01689 { 01690 if (first) 01691 { 01692 os << " out[" << i - firstOutputNeuron << "] = " << "z" << j 01693 << " * " << weightMatrix(i, j); 01694 first = false; 01695 } 01696 else 01697 { 01698 os << " + z" << j << " * " << weightMatrix(i, j); 01699 } 01700 } 01701 if( connectionMatrix(i, bias) ) { 01702 if (first) 01703 { 01704 os << " out[" << i - firstOutputNeuron << "] = " 01705 << weightMatrix(i, bias); // bias 01706 first = false; 01707 } 01708 else 01709 { 01710 os << " + " << weightMatrix(i, bias); // bias 01711 } 01712 } 01713 01714 if (!first) 01715 { 01716 os << ";\n"; 01717 if((strlen(gOut) != 1) || (*gOut != '#')) { // skip for linear output neurons 01718 replaceInString(gOut, str, "#", "out[", i - firstOutputNeuron, "]"); 01719 os << " out[" << i - firstOutputNeuron << "] = " << str << ";\n"; 01720 } 01721 } 01722 } 01723 os << "};" << std::endl; 01724 } 01725 01726