Relintai/pmlpp
1
0
Fork 0
mirror of https://github.com/Relintai/pmlpp.git synced 2024-11-12 13:47:18 +01:00
Code Issues Packages Projects Releases Wiki Activity

MLPPSoftmaxNet initial cleanup.

Browse Source
This commit is contained in:
Relintai 2023-02-11 09:17:02 +01:00
parent 47155163b1
commit 1bb0cab99a
4 changed files with 349 additions and 165 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
mlpp/data/data.cpp
View File

@ -10,7 +10,7 @@
#include "core/os/file_access.h" #include "core/os/file_access.h"
#include "../lin_alg/lin_alg.h" #include "../lin_alg/lin_alg.h"
#include "../softmax_net/softmax_net.h" #include "../softmax_net/softmax_net_old.h"
#include "../stat/stat.h" #include "../stat/stat.h"
#include <algorithm> #include <algorithm>
@ -1008,11 +1008,11 @@ std::tuple<std::vector<std::vector<real_t>>, std::vector<std::string>> MLPPData:
outputSet.push_back(BOW[i]); outputSet.push_back(BOW[i]);
} }
MLPPLinAlg alg; MLPPLinAlg alg;
MLPPSoftmaxNet *model; MLPPSoftmaxNetOld *model;
if (type == "Skipgram") { if (type == "Skipgram") {
model = new MLPPSoftmaxNet(outputSet, inputSet, dimension); model = new MLPPSoftmaxNetOld(outputSet, inputSet, dimension);
} else { // else = CBOW. We maintain it is a default. } else { // else = CBOW. We maintain it is a default.
model = new MLPPSoftmaxNet(inputSet, outputSet, dimension); model = new MLPPSoftmaxNetOld(inputSet, outputSet, dimension);
} }
model->gradientDescent(learning_rate, max_epoch, 1); model->gradientDescent(learning_rate, max_epoch, 1);
@ -1074,11 +1074,11 @@ MLPPData::WordsToVecResult MLPPData::word_to_vec(std::vector<std::string> senten
outputSet.push_back(BOW[i]); outputSet.push_back(BOW[i]);
} }
MLPPLinAlg alg; MLPPLinAlg alg;
MLPPSoftmaxNet *model; MLPPSoftmaxNetOld *model;
if (type == "Skipgram") { if (type == "Skipgram") {
model = new MLPPSoftmaxNet(outputSet, inputSet, dimension); model = new MLPPSoftmaxNetOld(outputSet, inputSet, dimension);
} else { // else = CBOW. We maintain it is a default. } else { // else = CBOW. We maintain it is a default.
model = new MLPPSoftmaxNet(inputSet, outputSet, dimension); model = new MLPPSoftmaxNetOld(inputSet, outputSet, dimension);
} }
model->gradientDescent(learning_rate, max_epoch, false); model->gradientDescent(learning_rate, max_epoch, false);

367
mlpp/softmax_net/softmax_net.cpp
View File

@ -15,81 +15,114 @@
#include <iostream> #include <iostream>
#include <random> #include <random>
MLPPSoftmaxNet::MLPPSoftmaxNet(std::vector<std::vector<real_t>> pinputSet, std::vector<std::vector<real_t>> poutputSet, int pn_hidden, std::string preg, real_t plambda, real_t palpha) { /*
inputSet = pinputSet; Ref<MLPPMatrix> MLPPSoftmaxNet::get_input_set() {
outputSet = poutputSet; return _input_set;
n = pinputSet.size(); }
k = pinputSet[0].size(); void MLPPSoftmaxNet::set_input_set(const Ref<MLPPMatrix> &val) {
n_hidden = pn_hidden; _input_set = val;
n_class = poutputSet[0].size();
reg = preg;
lambda = plambda;
alpha = palpha;
y_hat.resize(n); _initialized = false;
weights1 = MLPPUtilities::weightInitialization(k, n_hidden);
weights2 = MLPPUtilities::weightInitialization(n_hidden, n_class);
bias1 = MLPPUtilities::biasInitialization(n_hidden);
bias2 = MLPPUtilities::biasInitialization(n_class);
} }
std::vector<real_t> MLPPSoftmaxNet::modelTest(std::vector<real_t> x) { Ref<MLPPMatrix> MLPPSoftmaxNet::get_output_set() {
return Evaluate(x); return _output_set;
}
void MLPPSoftmaxNet::set_output_set(const Ref<MLPPMatrix> &val) {
_output_set = val;
_initialized = false;
} }
std::vector<std::vector<real_t>> MLPPSoftmaxNet::modelSetTest(std::vector<std::vector<real_t>> X) { MLPPReg::RegularizationType MLPPSoftmaxNet::get_reg() {
return Evaluate(X); return _reg;
}
void MLPPSoftmaxNet::set_reg(const MLPPReg::RegularizationType val) {
_reg = val;
_initialized = false;
} }
void MLPPSoftmaxNet::gradientDescent(real_t learning_rate, int max_epoch, bool UI) { real_t MLPPSoftmaxNet::get_lambda() {
return _lambda;
}
void MLPPSoftmaxNet::set_lambda(const real_t val) {
_lambda = val;
_initialized = false;
}
real_t MLPPSoftmaxNet::get_alpha() {
return _alpha;
}
void MLPPSoftmaxNet::set_alpha(const real_t val) {
_alpha = val;
_initialized = false;
}
*/
std::vector<real_t> MLPPSoftmaxNet::model_test(std::vector<real_t> x) {
return evaluatev(x);
}
std::vector<std::vector<real_t>> MLPPSoftmaxNet::model_set_test(std::vector<std::vector<real_t>> X) {
return evaluatem(X);
}
void MLPPSoftmaxNet::gradient_descent(real_t learning_rate, int max_epoch, bool ui) {
MLPPActivation avn; MLPPActivation avn;
MLPPLinAlg alg; MLPPLinAlg alg;
MLPPReg regularization; MLPPReg regularization;
real_t cost_prev = 0; real_t cost_prev = 0;
int epoch = 1; int epoch = 1;
forwardPass();
forward_pass();
while (true) { while (true) {
cost_prev = Cost(y_hat, outputSet); cost_prev = cost(_y_hat, _output_set);
// Calculating the errors // Calculating the errors
std::vector<std::vector<real_t>> error = alg.subtraction(y_hat, outputSet); std::vector<std::vector<real_t>> error = alg.subtraction(_y_hat, _output_set);
// Calculating the weight/bias gradients for layer 2 // Calculating the weight/bias gradients for layer 2
std::vector<std::vector<real_t>> D2_1 = alg.matmult(alg.transpose(a2), error); std::vector<std::vector<real_t>> D2_1 = alg.matmult(alg.transpose(_a2), error);
// weights and bias updation for layer 2 // weights and bias updation for layer 2
weights2 = alg.subtraction(weights2, alg.scalarMultiply(learning_rate, D2_1)); _weights2 = alg.subtraction(_weights2, alg.scalarMultiply(learning_rate, D2_1));
weights2 = regularization.regWeights(weights2, lambda, alpha, reg); //_reg
_weights2 = regularization.regWeights(_weights2, _lambda, _alpha, "None");
bias2 = alg.subtractMatrixRows(bias2, alg.scalarMultiply(learning_rate, error)); _bias2 = alg.subtractMatrixRows(_bias2, alg.scalarMultiply(learning_rate, error));
//Calculating the weight/bias for layer 1 //Calculating the weight/bias for layer 1
std::vector<std::vector<real_t>> D1_1 = alg.matmult(error, alg.transpose(weights2)); std::vector<std::vector<real_t>> D1_1 = alg.matmult(error, alg.transpose(_weights2));
std::vector<std::vector<real_t>> D1_2 = alg.hadamard_product(D1_1, avn.sigmoid(z2, 1)); std::vector<std::vector<real_t>> D1_2 = alg.hadamard_product(D1_1, avn.sigmoid(_z2, true));
std::vector<std::vector<real_t>> D1_3 = alg.matmult(alg.transpose(inputSet), D1_2); std::vector<std::vector<real_t>> D1_3 = alg.matmult(alg.transpose(_input_set), D1_2);
// weight an bias updation for layer 1 // weight an bias updation for layer 1
weights1 = alg.subtraction(weights1, alg.scalarMultiply(learning_rate, D1_3)); _weights1 = alg.subtraction(_weights1, alg.scalarMultiply(learning_rate, D1_3));
weights1 = regularization.regWeights(weights1, lambda, alpha, reg); //_reg
_weights1 = regularization.regWeights(_weights1, _lambda, _alpha, "None");
bias1 = alg.subtractMatrixRows(bias1, alg.scalarMultiply(learning_rate, D1_2)); _bias1 = alg.subtractMatrixRows(_bias1, alg.scalarMultiply(learning_rate, D1_2));
forwardPass(); forward_pass();
// UI PORTION // UI PORTION
if (UI) { if (ui) {
MLPPUtilities::CostInfo(epoch, cost_prev, Cost(y_hat, outputSet)); MLPPUtilities::CostInfo(epoch, cost_prev, cost(_y_hat, _output_set));
std::cout << "Layer 1:" << std::endl; std::cout << "Layer 1:" << std::endl;
MLPPUtilities::UI(weights1, bias1); MLPPUtilities::UI(_weights1, _bias1);
std::cout << "Layer 2:" << std::endl; std::cout << "Layer 2:" << std::endl;
MLPPUtilities::UI(weights2, bias2); MLPPUtilities::UI(_weights2, _bias2);
} }
epoch++; epoch++;
if (epoch > max_epoch) { if (epoch > max_epoch) {
@ -98,65 +131,72 @@ void MLPPSoftmaxNet::gradientDescent(real_t learning_rate, int max_epoch, bool U
} }
} }
void MLPPSoftmaxNet::SGD(real_t learning_rate, int max_epoch, bool UI) { void MLPPSoftmaxNet::sgd(real_t learning_rate, int max_epoch, bool ui) {
MLPPActivation avn; MLPPActivation avn;
MLPPLinAlg alg; MLPPLinAlg alg;
MLPPReg regularization; MLPPReg regularization;
real_t cost_prev = 0; real_t cost_prev = 0;
int epoch = 1; int epoch = 1;
std::random_device rd;
std::default_random_engine generator(rd());
std::uniform_int_distribution<int> distribution(0, int(_n - 1));
while (true) { while (true) {
std::random_device rd;
std::default_random_engine generator(rd());
std::uniform_int_distribution<int> distribution(0, int(n - 1));
int outputIndex = distribution(generator); int outputIndex = distribution(generator);
std::vector<real_t> y_hat = Evaluate(inputSet[outputIndex]); std::vector<real_t> y_hat = evaluatev(_input_set[outputIndex]);
auto prop_res = propagate(inputSet[outputIndex]); auto prop_res = propagatev(_input_set[outputIndex]);
auto z2 = std::get<0>(prop_res); auto z2 = std::get<0>(prop_res);
auto a2 = std::get<1>(prop_res); auto a2 = std::get<1>(prop_res);
cost_prev = Cost({ y_hat }, { outputSet[outputIndex] }); cost_prev = cost({ y_hat }, { _output_set[outputIndex] });
std::vector<real_t> error = alg.subtraction(y_hat, outputSet[outputIndex]); std::vector<real_t> error = alg.subtraction(y_hat, _output_set[outputIndex]);
// Weight updation for layer 2 // Weight updation for layer 2
std::vector<std::vector<real_t>> D2_1 = alg.outerProduct(error, a2); std::vector<std::vector<real_t>> D2_1 = alg.outerProduct(error, a2);
weights2 = alg.subtraction(weights2, alg.scalarMultiply(learning_rate, alg.transpose(D2_1))); _weights2 = alg.subtraction(_weights2, alg.scalarMultiply(learning_rate, alg.transpose(D2_1)));
weights2 = regularization.regWeights(weights2, lambda, alpha, reg); //_reg
_weights2 = regularization.regWeights(_weights2, _lambda, _alpha, "None");
// Bias updation for layer 2 // Bias updation for layer 2
bias2 = alg.subtraction(bias2, alg.scalarMultiply(learning_rate, error)); _bias2 = alg.subtraction(_bias2, alg.scalarMultiply(learning_rate, error));
// Weight updation for layer 1 // Weight updation for layer 1
std::vector<real_t> D1_1 = alg.mat_vec_mult(weights2, error); std::vector<real_t> D1_1 = alg.mat_vec_mult(_weights2, error);
std::vector<real_t> D1_2 = alg.hadamard_product(D1_1, avn.sigmoid(z2, true)); std::vector<real_t> D1_2 = alg.hadamard_product(D1_1, avn.sigmoid(z2, true));
std::vector<std::vector<real_t>> D1_3 = alg.outerProduct(inputSet[outputIndex], D1_2); std::vector<std::vector<real_t>> D1_3 = alg.outerProduct(_input_set[outputIndex], D1_2);
weights1 = alg.subtraction(weights1, alg.scalarMultiply(learning_rate, D1_3)); _weights1 = alg.subtraction(_weights1, alg.scalarMultiply(learning_rate, D1_3));
weights1 = regularization.regWeights(weights1, lambda, alpha, reg); //_reg
_weights1 = regularization.regWeights(_weights1, _lambda, _alpha, "None");
// Bias updation for layer 1 // Bias updation for layer 1
bias1 = alg.subtraction(bias1, alg.scalarMultiply(learning_rate, D1_2)); _bias1 = alg.subtraction(_bias1, alg.scalarMultiply(learning_rate, D1_2));
y_hat = Evaluate(inputSet[outputIndex]); y_hat = evaluatev(_input_set[outputIndex]);
if (UI) {
MLPPUtilities::CostInfo(epoch, cost_prev, Cost({ y_hat }, { outputSet[outputIndex] })); if (ui) {
MLPPUtilities::CostInfo(epoch, cost_prev, cost({ y_hat }, { _output_set[outputIndex] }));
std::cout << "Layer 1:" << std::endl; std::cout << "Layer 1:" << std::endl;
MLPPUtilities::UI(weights1, bias1); MLPPUtilities::UI(_weights1, _bias1);
std::cout << "Layer 2:" << std::endl; std::cout << "Layer 2:" << std::endl;
MLPPUtilities::UI(weights2, bias2); MLPPUtilities::UI(_weights2, _bias2);
} }
epoch++; epoch++;
if (epoch > max_epoch) { if (epoch > max_epoch) {
break; break;
} }
} }
forwardPass();
forward_pass();
} }
void MLPPSoftmaxNet::MBGD(real_t learning_rate, int max_epoch, int mini_batch_size, bool UI) { void MLPPSoftmaxNet::mbgd(real_t learning_rate, int max_epoch, int mini_batch_size, bool ui) {
MLPPActivation avn; MLPPActivation avn;
MLPPLinAlg alg; MLPPLinAlg alg;
MLPPReg regularization; MLPPReg regularization;
@ -164,40 +204,21 @@ void MLPPSoftmaxNet::MBGD(real_t learning_rate, int max_epoch, int mini_batch_si
int epoch = 1; int epoch = 1;
// Creating the mini-batches // Creating the mini-batches
int n_mini_batch = n / mini_batch_size; int n_mini_batch = _n / mini_batch_size;
auto batches = MLPPUtilities::createMiniBatches(inputSet, outputSet, n_mini_batch); auto batches = MLPPUtilities::createMiniBatches(_input_set, _output_set, n_mini_batch);
auto inputMiniBatches = std::get<0>(batches); auto inputMiniBatches = std::get<0>(batches);
auto outputMiniBatches = std::get<1>(batches); auto outputMiniBatches = std::get<1>(batches);
// Creating the mini-batches
for (int i = 0; i < n_mini_batch; i++) {
std::vector<std::vector<real_t>> currentInputSet;
std::vector<std::vector<real_t>> currentOutputSet;
for (int j = 0; j < n / n_mini_batch; j++) {
currentInputSet.push_back(inputSet[n / n_mini_batch * i + j]);
currentOutputSet.push_back(outputSet[n / n_mini_batch * i + j]);
}
inputMiniBatches.push_back(currentInputSet);
outputMiniBatches.push_back(currentOutputSet);
}
if (real_t(n) / real_t(n_mini_batch) - int(n / n_mini_batch) != 0) {
for (int i = 0; i < n - n / n_mini_batch * n_mini_batch; i++) {
inputMiniBatches[n_mini_batch - 1].push_back(inputSet[n / n_mini_batch * n_mini_batch + i]);
outputMiniBatches[n_mini_batch - 1].push_back(outputSet[n / n_mini_batch * n_mini_batch + i]);
}
}
while (true) { while (true) {
for (int i = 0; i < n_mini_batch; i++) { for (int i = 0; i < n_mini_batch; i++) {
std::vector<std::vector<real_t>> y_hat = Evaluate(inputMiniBatches[i]); std::vector<std::vector<real_t>> y_hat = evaluatem(inputMiniBatches[i]);
auto propagate_res = propagate(inputMiniBatches[i]); auto propagate_res = propagatem(inputMiniBatches[i]);
auto z2 = std::get<0>(propagate_res); auto z2 = std::get<0>(propagate_res);
auto a2 = std::get<1>(propagate_res); auto a2 = std::get<1>(propagate_res);
cost_prev = Cost(y_hat, outputMiniBatches[i]); cost_prev = cost(y_hat, outputMiniBatches[i]);
// Calculating the errors // Calculating the errors
std::vector<std::vector<real_t>> error = alg.subtraction(y_hat, outputMiniBatches[i]); std::vector<std::vector<real_t>> error = alg.subtraction(y_hat, outputMiniBatches[i]);
@ -207,102 +228,198 @@ void MLPPSoftmaxNet::MBGD(real_t learning_rate, int max_epoch, int mini_batch_si
std::vector<std::vector<real_t>> D2_1 = alg.matmult(alg.transpose(a2), error); std::vector<std::vector<real_t>> D2_1 = alg.matmult(alg.transpose(a2), error);
// weights and bias updation for layser 2 // weights and bias updation for layser 2
weights2 = alg.subtraction(weights2, alg.scalarMultiply(learning_rate, D2_1)); _weights2 = alg.subtraction(_weights2, alg.scalarMultiply(learning_rate, D2_1));
weights2 = regularization.regWeights(weights2, lambda, alpha, reg); //_reg
_weights2 = regularization.regWeights(_weights2, _lambda, _alpha, "None");
// Bias Updation for layer 2 // Bias Updation for layer 2
bias2 = alg.subtractMatrixRows(bias2, alg.scalarMultiply(learning_rate, error)); _bias2 = alg.subtractMatrixRows(_bias2, alg.scalarMultiply(learning_rate, error));
//Calculating the weight/bias for layer 1 //Calculating the weight/bias for layer 1
std::vector<std::vector<real_t>> D1_1 = alg.matmult(error, alg.transpose(weights2)); std::vector<std::vector<real_t>> D1_1 = alg.matmult(error, alg.transpose(_weights2));
std::vector<std::vector<real_t>> D1_2 = alg.hadamard_product(D1_1, avn.sigmoid(z2, 1)); std::vector<std::vector<real_t>> D1_2 = alg.hadamard_product(D1_1, avn.sigmoid(z2, true));
std::vector<std::vector<real_t>> D1_3 = alg.matmult(alg.transpose(inputMiniBatches[i]), D1_2); std::vector<std::vector<real_t>> D1_3 = alg.matmult(alg.transpose(inputMiniBatches[i]), D1_2);
// weight an bias updation for layer 1 // weight an bias updation for layer 1
weights1 = alg.subtraction(weights1, alg.scalarMultiply(learning_rate, D1_3)); _weights1 = alg.subtraction(_weights1, alg.scalarMultiply(learning_rate, D1_3));
weights1 = regularization.regWeights(weights1, lambda, alpha, reg); //_reg
_weights1 = regularization.regWeights(_weights1, _lambda, _alpha, "None");
bias1 = alg.subtractMatrixRows(bias1, alg.scalarMultiply(learning_rate, D1_2)); _bias1 = alg.subtractMatrixRows(_bias1, alg.scalarMultiply(learning_rate, D1_2));
y_hat = Evaluate(inputMiniBatches[i]); y_hat = evaluatem(inputMiniBatches[i]);
if (UI) { if (ui) {
MLPPUtilities::CostInfo(epoch, cost_prev, Cost(y_hat, outputMiniBatches[i])); MLPPUtilities::CostInfo(epoch, cost_prev, cost(y_hat, outputMiniBatches[i]));
std::cout << "Layer 1:" << std::endl; std::cout << "Layer 1:" << std::endl;
MLPPUtilities::UI(weights1, bias1); MLPPUtilities::UI(_weights1, _bias1);
std::cout << "Layer 2:" << std::endl; std::cout << "Layer 2:" << std::endl;
MLPPUtilities::UI(weights2, bias2); MLPPUtilities::UI(_weights2, _bias2);
} }
} }
epoch++; epoch++;
if (epoch > max_epoch) { if (epoch > max_epoch) {
break; break;
} }
} }
forwardPass();
forward_pass();
} }
real_t MLPPSoftmaxNet::score() { real_t MLPPSoftmaxNet::score() {
MLPPUtilities util; MLPPUtilities util;
return util.performance(y_hat, outputSet);
return util.performance(_y_hat, _output_set);
} }
void MLPPSoftmaxNet::save(std::string fileName) { void MLPPSoftmaxNet::save(std::string fileName) {
MLPPUtilities util; MLPPUtilities util;
util.saveParameters(fileName, weights1, bias1, 0, 1);
util.saveParameters(fileName, weights2, bias2, 1, 2); util.saveParameters(fileName, _weights1, _bias1, false, 1);
util.saveParameters(fileName, _weights2, _bias2, true, 2);
} }
std::vector<std::vector<real_t>> MLPPSoftmaxNet::getEmbeddings() { std::vector<std::vector<real_t>> MLPPSoftmaxNet::get_embeddings() {
return weights1; return _weights1;
} }
real_t MLPPSoftmaxNet::Cost(std::vector<std::vector<real_t>> y_hat, std::vector<std::vector<real_t>> y) { bool MLPPSoftmaxNet::is_initialized() {
return _initialized;
}
void MLPPSoftmaxNet::initialize() {
if (_initialized) {
return;
}
//ERR_FAIL_COND(!_input_set.is_valid() || !_output_set.is_valid());
_initialized = true;
}
MLPPSoftmaxNet::MLPPSoftmaxNet(std::vector<std::vector<real_t>> p_input_set, std::vector<std::vector<real_t>> p_output_set, int p_n_hidden, MLPPReg::RegularizationType p_reg, real_t p_lambda, real_t p_alpha) {
_input_set = p_input_set;
_output_set = p_output_set;
_n = p_input_set.size();
_k = p_input_set[0].size();
_n_hidden = p_n_hidden;
_n_class = p_output_set[0].size();
_reg = p_reg;
_lambda = p_lambda;
_alpha = p_alpha;
_y_hat.resize(_n);
_weights1 = MLPPUtilities::weightInitialization(_k, _n_hidden);
_weights2 = MLPPUtilities::weightInitialization(_n_hidden, _n_class);
_bias1 = MLPPUtilities::biasInitialization(_n_hidden);
_bias2 = MLPPUtilities::biasInitialization(_n_class);
_initialized = true;
}
MLPPSoftmaxNet::MLPPSoftmaxNet() {
_initialized = false;
}
MLPPSoftmaxNet::~MLPPSoftmaxNet() {
}
real_t MLPPSoftmaxNet::cost(std::vector<std::vector<real_t>> y_hat, std::vector<std::vector<real_t>> y) {
MLPPReg regularization; MLPPReg regularization;
MLPPData data; MLPPData data;
class MLPPCost cost; class MLPPCost cost;
return cost.CrossEntropy(y_hat, y) + regularization.regTerm(weights1, lambda, alpha, reg) + regularization.regTerm(weights2, lambda, alpha, reg);
//_reg
return cost.CrossEntropy(y_hat, y) + regularization.regTerm(_weights1, _lambda, _alpha, "None") + regularization.regTerm(_weights2, _lambda, _alpha, "None");
} }
std::vector<std::vector<real_t>> MLPPSoftmaxNet::Evaluate(std::vector<std::vector<real_t>> X) { std::vector<real_t> MLPPSoftmaxNet::evaluatev(std::vector<real_t> x) {
MLPPLinAlg alg; MLPPLinAlg alg;
MLPPActivation avn; MLPPActivation avn;
std::vector<std::vector<real_t>> z2 = alg.mat_vec_add(alg.matmult(X, weights1), bias1);
std::vector<std::vector<real_t>> a2 = avn.sigmoid(z2); std::vector<real_t> z2 = alg.addition(alg.mat_vec_mult(alg.transpose(_weights1), x), _bias1);
return avn.adjSoftmax(alg.mat_vec_add(alg.matmult(a2, weights2), bias2)); std::vector<real_t> a2 = avn.sigmoid(z2);
return avn.adjSoftmax(alg.addition(alg.mat_vec_mult(alg.transpose(_weights2), a2), _bias2));
} }
std::tuple<std::vector<std::vector<real_t>>, std::vector<std::vector<real_t>>> MLPPSoftmaxNet::propagate(std::vector<std::vector<real_t>> X) { std::tuple<std::vector<real_t>, std::vector<real_t>> MLPPSoftmaxNet::propagatev(std::vector<real_t> x) {
MLPPLinAlg alg; MLPPLinAlg alg;
MLPPActivation avn; MLPPActivation avn;
std::vector<std::vector<real_t>> z2 = alg.mat_vec_add(alg.matmult(X, weights1), bias1);
std::vector<std::vector<real_t>> a2 = avn.sigmoid(z2); std::vector<real_t> z2 = alg.addition(alg.mat_vec_mult(alg.transpose(_weights1), x), _bias1);
std::vector<real_t> a2 = avn.sigmoid(z2);
return { z2, a2 }; return { z2, a2 };
} }
std::vector<real_t> MLPPSoftmaxNet::Evaluate(std::vector<real_t> x) { std::vector<std::vector<real_t>> MLPPSoftmaxNet::evaluatem(std::vector<std::vector<real_t>> X) {
MLPPLinAlg alg; MLPPLinAlg alg;
MLPPActivation avn; MLPPActivation avn;
std::vector<real_t> z2 = alg.addition(alg.mat_vec_mult(alg.transpose(weights1), x), bias1);
std::vector<real_t> a2 = avn.sigmoid(z2); std::vector<std::vector<real_t>> z2 = alg.mat_vec_add(alg.matmult(X, _weights1), _bias1);
return avn.adjSoftmax(alg.addition(alg.mat_vec_mult(alg.transpose(weights2), a2), bias2)); std::vector<std::vector<real_t>> a2 = avn.sigmoid(z2);
return avn.adjSoftmax(alg.mat_vec_add(alg.matmult(a2, _weights2), _bias2));
} }
std::tuple<std::vector<real_t>, std::vector<real_t>> MLPPSoftmaxNet::propagate(std::vector<real_t> x) { std::tuple<std::vector<std::vector<real_t>>, std::vector<std::vector<real_t>>> MLPPSoftmaxNet::propagatem(std::vector<std::vector<real_t>> X) {
MLPPLinAlg alg; MLPPLinAlg alg;
MLPPActivation avn; MLPPActivation avn;
std::vector<real_t> z2 = alg.addition(alg.mat_vec_mult(alg.transpose(weights1), x), bias1);
std::vector<real_t> a2 = avn.sigmoid(z2); std::vector<std::vector<real_t>> z2 = alg.mat_vec_add(alg.matmult(X, _weights1), _bias1);
std::vector<std::vector<real_t>> a2 = avn.sigmoid(z2);
return { z2, a2 }; return { z2, a2 };
} }
void MLPPSoftmaxNet::forwardPass() { void MLPPSoftmaxNet::forward_pass() {
MLPPLinAlg alg; MLPPLinAlg alg;
MLPPActivation avn; MLPPActivation avn;
z2 = alg.mat_vec_add(alg.matmult(inputSet, weights1), bias1);
a2 = avn.sigmoid(z2); _z2 = alg.mat_vec_add(alg.matmult(_input_set, _weights1), _bias1);
y_hat = avn.adjSoftmax(alg.mat_vec_add(alg.matmult(a2, weights2), bias2)); _a2 = avn.sigmoid(_z2);
_y_hat = avn.adjSoftmax(alg.mat_vec_add(alg.matmult(_a2, _weights2), _bias2));
}
void MLPPSoftmaxNet::_bind_methods() {
/*
ClassDB::bind_method(D_METHOD("get_input_set"), &MLPPSoftmaxNet::get_input_set);
ClassDB::bind_method(D_METHOD("set_input_set", "val"), &MLPPSoftmaxNet::set_input_set);
ADD_PROPERTY(PropertyInfo(Variant::OBJECT, "input_set", PROPERTY_HINT_RESOURCE_TYPE, "MLPPMatrix"), "set_input_set", "get_input_set");
ClassDB::bind_method(D_METHOD("get_output_set"), &MLPPSoftmaxNet::get_output_set);
ClassDB::bind_method(D_METHOD("set_output_set", "val"), &MLPPSoftmaxNet::set_output_set);
ADD_PROPERTY(PropertyInfo(Variant::OBJECT, "output_set", PROPERTY_HINT_RESOURCE_TYPE, "MLPPMatrix"), "set_output_set", "get_output_set");
ClassDB::bind_method(D_METHOD("get_reg"), &MLPPSoftmaxNet::get_reg);
ClassDB::bind_method(D_METHOD("set_reg", "val"), &MLPPSoftmaxNet::set_reg);
ADD_PROPERTY(PropertyInfo(Variant::INT, "reg"), "set_reg", "get_reg");
ClassDB::bind_method(D_METHOD("get_lambda"), &MLPPSoftmaxNet::get_lambda);
ClassDB::bind_method(D_METHOD("set_lambda", "val"), &MLPPSoftmaxNet::set_lambda);
ADD_PROPERTY(PropertyInfo(Variant::REAL, "lambda"), "set_lambda", "get_lambda");
ClassDB::bind_method(D_METHOD("get_alpha"), &MLPPSoftmaxNet::get_alpha);
ClassDB::bind_method(D_METHOD("set_alpha", "val"), &MLPPSoftmaxNet::set_alpha);
ADD_PROPERTY(PropertyInfo(Variant::REAL, "alpha"), "set_alpha", "get_alpha");
ClassDB::bind_method(D_METHOD("model_test", "x"), &MLPPSoftmaxNet::model_test);
ClassDB::bind_method(D_METHOD("model_set_test", "X"), &MLPPSoftmaxNet::model_set_test);
ClassDB::bind_method(D_METHOD("gradient_descent", "learning_rate", "max_epoch", "ui"), &MLPPSoftmaxNet::gradient_descent, false);
ClassDB::bind_method(D_METHOD("sgd", "learning_rate", "max_epoch", "ui"), &MLPPSoftmaxNet::sgd, false);
ClassDB::bind_method(D_METHOD("mbgd", "learning_rate", "max_epoch", "mini_batch_size", "ui"), &MLPPSoftmaxNet::mbgd, false);
ClassDB::bind_method(D_METHOD("score"), &MLPPSoftmaxNet::score);
ClassDB::bind_method(D_METHOD("save", "file_name"), &MLPPSoftmaxNet::save);
ClassDB::bind_method(D_METHOD("is_initialized"), &MLPPSoftmaxNet::is_initialized);
ClassDB::bind_method(D_METHOD("initialize"), &MLPPSoftmaxNet::initialize);
*/
} }

106
mlpp/softmax_net/softmax_net.h
View File

@ -9,52 +9,96 @@
#include "core/math/math_defs.h" #include "core/math/math_defs.h"
#include "core/object/reference.h"
#include "../lin_alg/mlpp_matrix.h"
#include "../lin_alg/mlpp_vector.h"
#include "../regularization/reg.h"
#include <string> #include <string>
#include <vector> #include <vector>
class MLPPSoftmaxNet { class MLPPSoftmaxNet : public Reference {
GDCLASS(MLPPSoftmaxNet, Reference);
public: public:
MLPPSoftmaxNet(std::vector<std::vector<real_t>> inputSet, std::vector<std::vector<real_t>> outputSet, int n_hidden, std::string reg = "None", real_t lambda = 0.5, real_t alpha = 0.5); /*
std::vector<real_t> modelTest(std::vector<real_t> x); Ref<MLPPMatrix> get_input_set();
std::vector<std::vector<real_t>> modelSetTest(std::vector<std::vector<real_t>> X); void set_input_set(const Ref<MLPPMatrix> &val);
void gradientDescent(real_t learning_rate, int max_epoch, bool UI = false);
void SGD(real_t learning_rate, int max_epoch, bool UI = false); Ref<MLPPMatrix> get_output_set();
void MBGD(real_t learning_rate, int max_epoch, int mini_batch_size, bool UI = false); void set_output_set(const Ref<MLPPMatrix> &val);
MLPPReg::RegularizationType get_reg();
void set_reg(const MLPPReg::RegularizationType val);
real_t get_lambda();
void set_lambda(const real_t val);
real_t get_alpha();
void set_alpha(const real_t val);
*/
std::vector<real_t> model_test(std::vector<real_t> x);
std::vector<std::vector<real_t>> model_set_test(std::vector<std::vector<real_t>> X);
void gradient_descent(real_t learning_rate, int max_epoch, bool ui = false);
void sgd(real_t learning_rate, int max_epoch, bool ui = false);
void mbgd(real_t learning_rate, int max_epoch, int mini_batch_size, bool ui = false);
real_t score(); real_t score();
void save(std::string fileName); void save(std::string fileName);
std::vector<std::vector<real_t>> getEmbeddings(); // This class is used (mostly) for word2Vec. This function returns our embeddings. std::vector<std::vector<real_t>> get_embeddings(); // This class is used (mostly) for word2Vec. This function returns our embeddings.
private:
real_t Cost(std::vector<std::vector<real_t>> y_hat, std::vector<std::vector<real_t>> y);
std::vector<std::vector<real_t>> Evaluate(std::vector<std::vector<real_t>> X); bool is_initialized();
std::tuple<std::vector<std::vector<real_t>>, std::vector<std::vector<real_t>>> propagate(std::vector<std::vector<real_t>> X); void initialize();
std::vector<real_t> Evaluate(std::vector<real_t> x);
std::tuple<std::vector<real_t>, std::vector<real_t>> propagate(std::vector<real_t> x);
void forwardPass();
std::vector<std::vector<real_t>> inputSet; MLPPSoftmaxNet(std::vector<std::vector<real_t>> p_input_set, std::vector<std::vector<real_t>> p_output_set, int p_n_hidden, MLPPReg::RegularizationType p_reg = MLPPReg::REGULARIZATION_TYPE_NONE, real_t p_lambda = 0.5, real_t p_alpha = 0.5);
std::vector<std::vector<real_t>> outputSet; //MLPPSoftmaxNet(const Ref<MLPPMatrix> &p_input_set, const Ref<MLPPMatrix> &p_output_set, MLPPReg::RegularizationType p_reg = MLPPReg::REGULARIZATION_TYPE_NONE, real_t p_lambda = 0.5, real_t p_alpha = 0.5);
std::vector<std::vector<real_t>> y_hat;
std::vector<std::vector<real_t>> weights1; MLPPSoftmaxNet();
std::vector<std::vector<real_t>> weights2; ~MLPPSoftmaxNet();
std::vector<real_t> bias1; protected:
std::vector<real_t> bias2; real_t cost(std::vector<std::vector<real_t>> y_hat, std::vector<std::vector<real_t>> y);
std::vector<std::vector<real_t>> z2; std::vector<real_t> evaluatev(std::vector<real_t> x);
std::vector<std::vector<real_t>> a2; std::tuple<std::vector<real_t>, std::vector<real_t>> propagatev(std::vector<real_t> x);
int n; std::vector<std::vector<real_t>> evaluatem(std::vector<std::vector<real_t>> X);
int k; std::tuple<std::vector<std::vector<real_t>>, std::vector<std::vector<real_t>>> propagatem(std::vector<std::vector<real_t>> X);
int n_class;
int n_hidden; void forward_pass();
static void _bind_methods();
std::vector<std::vector<real_t>> _input_set;
std::vector<std::vector<real_t>> _output_set;
std::vector<std::vector<real_t>> _y_hat;
std::vector<std::vector<real_t>> _weights1;
std::vector<std::vector<real_t>> _weights2;
std::vector<real_t> _bias1;
std::vector<real_t> _bias2;
std::vector<std::vector<real_t>> _z2;
std::vector<std::vector<real_t>> _a2;
int _n;
int _k;
int _n_class;
int _n_hidden;
// Regularization Params // Regularization Params
std::string reg; MLPPReg::RegularizationType _reg;
real_t lambda; real_t _lambda;
real_t alpha; /* This is the controlling param for Elastic Net*/ real_t _alpha; /* This is the controlling param for Elastic Net*/
bool _initialized;
}; };
#endif /* SoftmaxNet_hpp */ #endif /* SoftmaxNet_hpp */

27
test/mlpp_tests.cpp
View File

@ -52,12 +52,30 @@
#include "../mlpp/outlier_finder/outlier_finder_old.h" #include "../mlpp/outlier_finder/outlier_finder_old.h"
#include "../mlpp/pca/pca_old.h" #include "../mlpp/pca/pca_old.h"
#include "../mlpp/probit_reg/probit_reg_old.h" #include "../mlpp/probit_reg/probit_reg_old.h"
#include "../mlpp/softmax_net/softmax_net_old.h"
#include "../mlpp/softmax_reg/softmax_reg_old.h" #include "../mlpp/softmax_reg/softmax_reg_old.h"
#include "../mlpp/svc/svc_old.h" #include "../mlpp/svc/svc_old.h"
#include "../mlpp/tanh_reg/tanh_reg_old.h" #include "../mlpp/tanh_reg/tanh_reg_old.h"
#include "../mlpp/uni_lin_reg/uni_lin_reg_old.h" #include "../mlpp/uni_lin_reg/uni_lin_reg_old.h"
#include "../mlpp/wgan/wgan_old.h" #include "../mlpp/wgan/wgan_old.h"
/*
#include "../mlpp/ann/ann_old.h"
#include "../mlpp/bernoulli_nb/bernoulli_nb_old.h"
#include "../mlpp/c_log_log_reg/c_log_log_reg_old.h"
#include "../mlpp/dual_svc/dual_svc_old.h"
#include "../mlpp/exp_reg/exp_reg_old.h"
#include "../mlpp/gan/gan_old.h"
#include "../mlpp/gaussian_nb/gaussian_nb_old.h"
#include "../mlpp/hidden_layer/hidden_layer_old.h"
#include "../mlpp/lin_reg/lin_reg_old.h"
#include "../mlpp/log_reg/log_reg_old.h"
#include "../mlpp/mann/mann_old.h"
#include "../mlpp/multi_output_layer/multi_output_layer_old.h"
#include "../mlpp/multinomial_nb/multinomial_nb_old.h"
#include "../mlpp/output_layer/output_layer_old.h"
*/
Vector<real_t> dstd_vec_to_vec(const std::vector<real_t> &in) { Vector<real_t> dstd_vec_to_vec(const std::vector<real_t> &in) {
Vector<real_t> r; Vector<real_t> r;
@ -490,9 +508,14 @@ void MLPPTests::test_soft_max_network(bool ui) {
// SOFTMAX NETWORK // SOFTMAX NETWORK
Ref<MLPPDataComplex> dt = data.load_wine(_wine_data_path); Ref<MLPPDataComplex> dt = data.load_wine(_wine_data_path);
MLPPSoftmaxNetOld model_old(dt->get_input()->to_std_vector(), dt->get_output()->to_std_vector(), 1);
model_old.gradientDescent(0.01, 100000, ui);
alg.printMatrix(model_old.modelSetTest(dt->get_input()->to_std_vector()));
std::cout << "ACCURACY: " << 100 * model_old.score() << "%" << std::endl;
MLPPSoftmaxNet model(dt->get_input()->to_std_vector(), dt->get_output()->to_std_vector(), 1); MLPPSoftmaxNet model(dt->get_input()->to_std_vector(), dt->get_output()->to_std_vector(), 1);
model.gradientDescent(0.01, 100000, ui); model.gradient_descent(0.01, 100000, ui);
alg.printMatrix(model.modelSetTest(dt->get_input()->to_std_vector())); alg.printMatrix(model.model_set_test(dt->get_input()->to_std_vector()));
std::cout << "ACCURACY: " << 100 * model.score() << "%" << std::endl; std::cout << "ACCURACY: " << 100 * model.score() << "%" << std::endl;
} }
void MLPPTests::test_autoencoder(bool ui) { void MLPPTests::test_autoencoder(bool ui) {