pmlpp/mlpp/softmax_net/softmax_net.cpp

455 lines
13 KiB
C++
Raw Normal View History

//
// SoftmaxNet.cpp
//
// Created by Marc Melikyan on 10/2/20.
//
2023-01-24 18:12:23 +01:00
#include "softmax_net.h"
2023-01-24 19:00:54 +01:00
#include "../activation/activation.h"
#include "../cost/cost.h"
2023-01-24 18:12:23 +01:00
#include "../data/data.h"
2023-01-24 19:00:54 +01:00
#include "../lin_alg/lin_alg.h"
2023-01-24 18:12:23 +01:00
#include "../regularization/reg.h"
#include "../utilities/utilities.h"
#include "core/log/logger.h"
#include <random>
2023-02-11 09:17:02 +01:00
/*
Ref<MLPPMatrix> MLPPSoftmaxNet::get_input_set() {
return _input_set;
2023-01-24 19:00:54 +01:00
}
2023-02-11 09:17:02 +01:00
void MLPPSoftmaxNet::set_input_set(const Ref<MLPPMatrix> &val) {
_input_set = val;
2023-01-24 19:00:54 +01:00
2023-02-11 09:17:02 +01:00
_initialized = false;
2023-01-24 19:00:54 +01:00
}
2023-02-11 09:17:02 +01:00
Ref<MLPPMatrix> MLPPSoftmaxNet::get_output_set() {
return _output_set;
2023-01-24 19:00:54 +01:00
}
2023-02-11 09:17:02 +01:00
void MLPPSoftmaxNet::set_output_set(const Ref<MLPPMatrix> &val) {
_output_set = val;
2023-01-24 19:00:54 +01:00
2023-02-11 09:17:02 +01:00
_initialized = false;
}
MLPPReg::RegularizationType MLPPSoftmaxNet::get_reg() {
return _reg;
}
void MLPPSoftmaxNet::set_reg(const MLPPReg::RegularizationType val) {
_reg = val;
_initialized = false;
}
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;
}
*/
Ref<MLPPVector> MLPPSoftmaxNet::model_test(const Ref<MLPPVector> &x) {
2023-02-11 09:17:02 +01:00
return evaluatev(x);
}
Ref<MLPPMatrix> MLPPSoftmaxNet::model_set_test(const Ref<MLPPMatrix> &X) {
2023-02-11 09:17:02 +01:00
return evaluatem(X);
}
void MLPPSoftmaxNet::gradient_descent(real_t learning_rate, int max_epoch, bool ui) {
2023-01-24 19:23:30 +01:00
MLPPActivation avn;
2023-01-25 00:29:02 +01:00
MLPPLinAlg alg;
2023-01-25 00:54:50 +01:00
MLPPReg regularization;
2023-02-11 09:17:02 +01:00
2023-01-27 13:01:16 +01:00
real_t cost_prev = 0;
2023-01-24 19:00:54 +01:00
int epoch = 1;
2023-02-11 09:17:02 +01:00
forward_pass();
2023-01-24 19:00:54 +01:00
while (true) {
2023-02-11 09:17:02 +01:00
cost_prev = cost(_y_hat, _output_set);
2023-01-24 19:00:54 +01:00
// Calculating the errors
Ref<MLPPMatrix> error = alg.subtractionm(_y_hat, _output_set);
2023-01-24 19:00:54 +01:00
// Calculating the weight/bias gradients for layer 2
Ref<MLPPMatrix> D2_1 = alg.matmultm(alg.transposem(_a2), error);
2023-01-24 19:00:54 +01:00
// weights and bias updation for layer 2
_weights2 = alg.subtractionm(_weights2, alg.scalar_multiplym(learning_rate, D2_1));
_weights2 = regularization.reg_weightsm(_weights2, _lambda, _alpha, _reg);
2023-01-24 19:00:54 +01:00
_bias2 = alg.subtract_matrix_rows(_bias2, alg.scalar_multiplym(learning_rate, error));
2023-01-24 19:00:54 +01:00
//Calculating the weight/bias for layer 1
Ref<MLPPMatrix> D1_1 = alg.matmultm(error, alg.transposem(_weights2));
2023-01-24 19:00:54 +01:00
Ref<MLPPMatrix> D1_2 = alg.hadamard_productm(D1_1, avn.sigmoid_derivm(_z2));
2023-01-24 19:00:54 +01:00
Ref<MLPPMatrix> D1_3 = alg.matmultm(alg.transposem(_input_set), D1_2);
2023-01-24 19:00:54 +01:00
// weight an bias updation for layer 1
_weights1 = alg.subtractionm(_weights1, alg.scalar_multiplym(learning_rate, D1_3));
_weights1 = regularization.reg_weightsm(_weights1, _lambda, _alpha, _reg);
2023-01-24 19:00:54 +01:00
_bias1 = alg.subtract_matrix_rows(_bias1, alg.scalar_multiplym(learning_rate, D1_2));
2023-01-24 19:00:54 +01:00
2023-02-11 09:17:02 +01:00
forward_pass();
2023-01-24 19:00:54 +01:00
// UI PORTION
2023-02-11 09:17:02 +01:00
if (ui) {
MLPPUtilities::cost_info(epoch, cost_prev, cost(_y_hat, _output_set));
PLOG_MSG("Layer 1:");
MLPPUtilities::print_ui_mb(_weights1, _bias1);
PLOG_MSG("Layer 2:");
MLPPUtilities::print_ui_mb(_weights2, _bias2);
2023-01-24 19:00:54 +01:00
}
2023-02-11 09:17:02 +01:00
2023-01-24 19:00:54 +01:00
epoch++;
if (epoch > max_epoch) {
break;
}
}
}
2023-02-11 09:17:02 +01:00
void MLPPSoftmaxNet::sgd(real_t learning_rate, int max_epoch, bool ui) {
2023-01-24 19:23:30 +01:00
MLPPActivation avn;
2023-01-25 00:29:02 +01:00
MLPPLinAlg alg;
2023-01-25 00:54:50 +01:00
MLPPReg regularization;
2023-02-11 09:17:02 +01:00
2023-01-27 13:01:16 +01:00
real_t cost_prev = 0;
2023-01-24 19:00:54 +01:00
int epoch = 1;
2023-02-11 09:17:02 +01:00
std::random_device rd;
std::default_random_engine generator(rd());
std::uniform_int_distribution<int> distribution(0, int(_n - 1));
Ref<MLPPVector> input_set_row_tmp;
input_set_row_tmp.instance();
input_set_row_tmp->resize(_input_set->size().x);
Ref<MLPPVector> output_set_row_tmp;
output_set_row_tmp.instance();
output_set_row_tmp->resize(_output_set->size().x);
Ref<MLPPMatrix> y_hat_mat_tmp;
y_hat_mat_tmp.instance();
y_hat_mat_tmp->resize(Size2i(_bias1->size(), 1));
Ref<MLPPMatrix> output_row_mat_tmp;
output_row_mat_tmp.instance();
output_row_mat_tmp->resize(Size2i(_output_set->size().x, 1));
2023-01-24 19:00:54 +01:00
while (true) {
int output_index = distribution(generator);
2023-01-24 19:00:54 +01:00
_input_set->get_row_into_mlpp_vector(output_index, input_set_row_tmp);
_output_set->get_row_into_mlpp_vector(output_index, output_set_row_tmp);
output_row_mat_tmp->set_row_mlpp_vector(0, output_set_row_tmp);
2023-02-10 21:26:46 +01:00
Ref<MLPPVector> y_hat = evaluatev(input_set_row_tmp);
y_hat_mat_tmp->set_row_mlpp_vector(0, y_hat);
2023-02-10 21:26:46 +01:00
PropagateVResult prop_res = propagatev(input_set_row_tmp);
cost_prev = cost(y_hat_mat_tmp, output_row_mat_tmp);
Ref<MLPPVector> error = alg.subtractionnv(y_hat, output_set_row_tmp);
2023-01-24 19:00:54 +01:00
// Weight updation for layer 2
Ref<MLPPMatrix> D2_1 = alg.outer_product(error, prop_res.a2);
_weights2 = alg.subtractionm(_weights2, alg.scalar_multiplym(learning_rate, alg.transposem(D2_1)));
_weights2 = regularization.reg_weightsm(_weights2, _lambda, _alpha, _reg);
2023-01-24 19:00:54 +01:00
// Bias updation for layer 2
_bias2 = alg.subtractionnv(_bias2, alg.scalar_multiplynv(learning_rate, error));
2023-01-24 19:00:54 +01:00
// Weight updation for layer 1
Ref<MLPPVector> D1_1 = alg.mat_vec_multv(_weights2, error);
Ref<MLPPVector> D1_2 = alg.hadamard_productm(D1_1, avn.sigmoid_derivv(prop_res.z2));
Ref<MLPPMatrix> D1_3 = alg.outer_product(input_set_row_tmp, D1_2);
2023-01-24 19:00:54 +01:00
_weights1 = alg.subtractionm(_weights1, alg.scalar_multiplym(learning_rate, D1_3));
_weights1 = regularization.reg_weightsm(_weights1, _lambda, _alpha, _reg);
2023-01-24 19:00:54 +01:00
// Bias updation for layer 1
_bias1 = alg.subtractionnv(_bias1, alg.scalar_multiplynv(learning_rate, D1_2));
2023-01-24 19:00:54 +01:00
y_hat = evaluatev(input_set_row_tmp);
2023-02-11 09:17:02 +01:00
if (ui) {
MLPPUtilities::cost_info(epoch, cost_prev, cost(y_hat_mat_tmp, output_row_mat_tmp));
PLOG_MSG("Layer 1:");
MLPPUtilities::print_ui_mb(_weights1, _bias1);
PLOG_MSG("Layer 2:");
MLPPUtilities::print_ui_mb(_weights2, _bias2);
2023-01-24 19:00:54 +01:00
}
2023-02-11 09:17:02 +01:00
2023-01-24 19:00:54 +01:00
epoch++;
if (epoch > max_epoch) {
break;
}
}
2023-02-11 09:17:02 +01:00
forward_pass();
2023-01-24 19:00:54 +01:00
}
2023-02-11 09:17:02 +01:00
void MLPPSoftmaxNet::mbgd(real_t learning_rate, int max_epoch, int mini_batch_size, bool ui) {
2023-01-24 19:23:30 +01:00
MLPPActivation avn;
2023-01-25 00:29:02 +01:00
MLPPLinAlg alg;
2023-01-25 00:54:50 +01:00
MLPPReg regularization;
2023-01-27 13:01:16 +01:00
real_t cost_prev = 0;
2023-01-24 19:00:54 +01:00
int epoch = 1;
// Creating the mini-batches
2023-02-11 09:17:02 +01:00
int n_mini_batch = _n / mini_batch_size;
2023-02-10 21:26:46 +01:00
MLPPUtilities::CreateMiniBatchMMBatch batches = MLPPUtilities::create_mini_batchesmm(_input_set, _output_set, n_mini_batch);
2023-01-24 19:00:54 +01:00
while (true) {
for (int i = 0; i < n_mini_batch; i++) {
Ref<MLPPMatrix> current_input_mini_batch = batches.input_sets[i];
Ref<MLPPMatrix> current_output_mini_batch = batches.output_sets[i];
Ref<MLPPMatrix> y_hat = evaluatem(current_input_mini_batch);
2023-02-10 21:26:46 +01:00
PropagateMResult prop_res = propagatem(current_input_mini_batch);
2023-02-10 21:26:46 +01:00
cost_prev = cost(y_hat, current_output_mini_batch);
2023-01-24 19:00:54 +01:00
// Calculating the errors
Ref<MLPPMatrix> error = alg.subtractionm(y_hat, current_output_mini_batch);
2023-01-24 19:00:54 +01:00
// Calculating the weight/bias gradients for layer 2
Ref<MLPPMatrix> D2_1 = alg.matmultm(alg.transposem(prop_res.a2), error);
2023-01-24 19:00:54 +01:00
// weights and bias updation for layser 2
_weights2 = alg.subtractionm(_weights2, alg.scalar_multiplym(learning_rate, D2_1));
_weights2 = regularization.reg_weightsm(_weights2, _lambda, _alpha, _reg);
2023-01-24 19:00:54 +01:00
// Bias Updation for layer 2
_bias2 = alg.subtract_matrix_rows(_bias2, alg.scalar_multiplym(learning_rate, error));
2023-01-24 19:00:54 +01:00
//Calculating the weight/bias for layer 1
Ref<MLPPMatrix> D1_1 = alg.matmultm(error, alg.transposem(_weights2));
Ref<MLPPMatrix> D1_2 = alg.hadamard_productm(D1_1, avn.sigmoid_derivm(prop_res.z2));
Ref<MLPPMatrix> D1_3 = alg.matmultm(alg.transposem(current_input_mini_batch), D1_2);
2023-01-24 19:00:54 +01:00
// weight an bias updation for layer 1
_weights1 = alg.subtractionm(_weights1, alg.scalar_multiplym(learning_rate, D1_3));
_weights1 = regularization.reg_weightsm(_weights1, _lambda, _alpha, _reg);
2023-01-24 19:00:54 +01:00
_bias1 = alg.subtract_matrix_rows(_bias1, alg.scalar_multiplym(learning_rate, D1_2));
2023-01-24 19:00:54 +01:00
y_hat = evaluatem(current_input_mini_batch);
2023-01-24 19:00:54 +01:00
2023-02-11 09:17:02 +01:00
if (ui) {
MLPPUtilities::cost_info(epoch, cost_prev, cost(y_hat, current_output_mini_batch));
PLOG_MSG("Layer 1:");
MLPPUtilities::print_ui_mb(_weights1, _bias1);
PLOG_MSG("Layer 2:");
MLPPUtilities::print_ui_mb(_weights2, _bias2);
2023-01-24 19:00:54 +01:00
}
}
2023-02-11 09:17:02 +01:00
2023-01-24 19:00:54 +01:00
epoch++;
2023-02-11 09:17:02 +01:00
2023-01-24 19:00:54 +01:00
if (epoch > max_epoch) {
break;
}
}
2023-02-11 09:17:02 +01:00
forward_pass();
2023-01-24 19:00:54 +01:00
}
2023-01-27 13:01:16 +01:00
real_t MLPPSoftmaxNet::score() {
2023-02-10 21:26:46 +01:00
MLPPUtilities util;
2023-02-11 09:17:02 +01:00
return util.performance_mat(_y_hat, _output_set);
2023-01-24 19:00:54 +01:00
}
void MLPPSoftmaxNet::save(const String &file_name) {
2023-02-10 21:26:46 +01:00
MLPPUtilities util;
2023-02-11 09:17:02 +01:00
//util.saveParameters(fileName, _weights1, _bias1, false, 1);
//util.saveParameters(fileName, _weights2, _bias2, true, 2);
2023-02-11 09:17:02 +01:00
}
Ref<MLPPMatrix> MLPPSoftmaxNet::get_embeddings() {
2023-02-11 09:17:02 +01:00
return _weights1;
}
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;
2023-01-24 19:00:54 +01:00
}
MLPPSoftmaxNet::MLPPSoftmaxNet(const Ref<MLPPMatrix> &p_input_set, const Ref<MLPPMatrix> &p_output_set, int p_n_hidden, MLPPReg::RegularizationType p_reg, real_t p_lambda, real_t p_alpha) {
2023-02-11 09:17:02 +01:00
_input_set = p_input_set;
_output_set = p_output_set;
_n = p_input_set->size().y;
_k = p_input_set->size().x;
2023-02-11 09:17:02 +01:00
_n_hidden = p_n_hidden;
_n_class = p_output_set->size().x;
2023-02-11 09:17:02 +01:00
_reg = p_reg;
_lambda = p_lambda;
_alpha = p_alpha;
_y_hat.instance();
_y_hat->resize(Size2i(0, _n));
MLPPUtilities utils;
_weights1.instance();
_weights1->resize(Size2i(_n_hidden, _k));
utils.weight_initializationm(_weights1);
2023-02-11 09:17:02 +01:00
_weights2.instance();
_weights2->resize(Size2i(_n_class, _n_hidden));
utils.weight_initializationm(_weights2);
_bias1.instance();
_bias1->resize(_n_hidden);
utils.bias_initializationv(_bias1);
_bias2.instance();
_bias2->resize(_n_class);
utils.bias_initializationv(_bias2);
2023-02-11 09:17:02 +01:00
_initialized = true;
}
MLPPSoftmaxNet::MLPPSoftmaxNet() {
_initialized = false;
}
MLPPSoftmaxNet::~MLPPSoftmaxNet() {
2023-01-24 19:00:54 +01:00
}
real_t MLPPSoftmaxNet::cost(const Ref<MLPPMatrix> &y_hat, const Ref<MLPPMatrix> &y) {
2023-01-25 00:54:50 +01:00
MLPPReg regularization;
2023-01-25 00:21:31 +01:00
MLPPData data;
MLPPCost mlpp_cost;
2023-02-11 09:17:02 +01:00
return mlpp_cost.cross_entropym(y_hat, y) + regularization.reg_termm(_weights1, _lambda, _alpha, _reg) + regularization.reg_termm(_weights2, _lambda, _alpha, _reg);
2023-01-24 19:00:54 +01:00
}
Ref<MLPPVector> MLPPSoftmaxNet::evaluatev(const Ref<MLPPVector> &x) {
2023-01-25 00:29:02 +01:00
MLPPLinAlg alg;
2023-01-24 19:23:30 +01:00
MLPPActivation avn;
2023-02-11 09:17:02 +01:00
Ref<MLPPVector> z2 = alg.additionnv(alg.mat_vec_multv(alg.transposem(_weights1), x), _bias1);
Ref<MLPPVector> a2 = avn.sigmoid_normv(z2);
2023-02-11 09:17:02 +01:00
return avn.adj_softmax_normv(alg.additionnv(alg.mat_vec_multv(alg.transposem(_weights2), a2), _bias2));
2023-01-24 19:00:54 +01:00
}
MLPPSoftmaxNet::PropagateVResult MLPPSoftmaxNet::propagatev(const Ref<MLPPVector> &x) {
2023-01-25 00:29:02 +01:00
MLPPLinAlg alg;
2023-01-24 19:23:30 +01:00
MLPPActivation avn;
2023-02-11 09:17:02 +01:00
PropagateVResult res;
2023-02-11 09:17:02 +01:00
res.z2 = alg.additionnv(alg.mat_vec_multv(alg.transposem(_weights1), x), _bias1);
res.a2 = avn.sigmoid_normv(res.z2);
return res;
2023-01-24 19:00:54 +01:00
}
Ref<MLPPMatrix> MLPPSoftmaxNet::evaluatem(const Ref<MLPPMatrix> &X) {
2023-01-25 00:29:02 +01:00
MLPPLinAlg alg;
2023-01-24 19:23:30 +01:00
MLPPActivation avn;
2023-02-11 09:17:02 +01:00
Ref<MLPPMatrix> z2 = alg.mat_vec_addv(alg.matmultm(X, _weights1), _bias1);
Ref<MLPPMatrix> a2 = avn.sigmoid_normm(z2);
2023-02-11 09:17:02 +01:00
return avn.adj_softmax_normm(alg.mat_vec_addv(alg.matmultm(a2, _weights2), _bias2));
2023-01-24 19:00:54 +01:00
}
MLPPSoftmaxNet::PropagateMResult MLPPSoftmaxNet::propagatem(const Ref<MLPPMatrix> &X) {
2023-01-25 00:29:02 +01:00
MLPPLinAlg alg;
2023-01-24 19:23:30 +01:00
MLPPActivation avn;
2023-02-11 09:17:02 +01:00
MLPPSoftmaxNet::PropagateMResult res;
res.z2 = alg.mat_vec_addv(alg.matmultm(X, _weights1), _bias1);
res.a2 = avn.sigmoid_normm(res.z2);
2023-02-11 09:17:02 +01:00
return res;
2023-01-24 19:00:54 +01:00
}
2023-02-11 09:17:02 +01:00
void MLPPSoftmaxNet::forward_pass() {
2023-01-25 00:29:02 +01:00
MLPPLinAlg alg;
2023-01-24 19:23:30 +01:00
MLPPActivation avn;
2023-02-11 09:17:02 +01:00
_z2 = alg.mat_vec_addv(alg.matmultm(_input_set, _weights1), _bias1);
_a2 = avn.sigmoid_normm(_z2);
_y_hat = avn.adj_softmax_normm(alg.mat_vec_addv(alg.matmultm(_a2, _weights2), _bias2));
2023-02-11 09:17:02 +01:00
}
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);
*/
2023-01-24 19:00:54 +01:00
}