mirror of
https://github.com/Relintai/pmlpp.git
synced 2024-12-23 15:06:52 +01:00
820 lines
26 KiB
C++
820 lines
26 KiB
C++
//
|
|
// LinReg.cpp
|
|
//
|
|
// Created by Marc Melikyan on 10/2/20.
|
|
//
|
|
|
|
#include "lin_reg.h"
|
|
|
|
#include "../cost/cost.h"
|
|
#include "../lin_alg/lin_alg.h"
|
|
#include "../regularization/reg.h"
|
|
#include "../stat/stat.h"
|
|
#include "../utilities/utilities.h"
|
|
|
|
#include <cmath>
|
|
#include <iostream>
|
|
#include <random>
|
|
|
|
/*
|
|
Ref<MLPPMatrix> MLPPLinReg::get_input_set() {
|
|
return _input_set;
|
|
}
|
|
void MLPPLinReg::set_input_set(const Ref<MLPPMatrix> &val) {
|
|
_input_set = val;
|
|
|
|
_initialized = false;
|
|
}
|
|
|
|
Ref<MLPPVector> MLPPLinReg::get_output_set() {
|
|
return _output_set;
|
|
}
|
|
void MLPPLinReg::set_output_set(const Ref<MLPPVector> &val) {
|
|
_output_set = val;
|
|
|
|
_initialized = false;
|
|
}
|
|
|
|
MLPPReg::RegularizationType MLPPLinReg::get_reg() {
|
|
return _reg;
|
|
}
|
|
void MLPPLinReg::set_reg(const MLPPReg::RegularizationType val) {
|
|
_reg = val;
|
|
|
|
_initialized = false;
|
|
}
|
|
|
|
real_t MLPPLinReg::get_lambda() {
|
|
return _lambda;
|
|
}
|
|
void MLPPLinReg::set_lambda(const real_t val) {
|
|
_lambda = val;
|
|
|
|
_initialized = false;
|
|
}
|
|
|
|
real_t MLPPLinReg::get_alpha() {
|
|
return _alpha;
|
|
}
|
|
void MLPPLinReg::set_alpha(const real_t val) {
|
|
_alpha = val;
|
|
|
|
_initialized = false;
|
|
}
|
|
*/
|
|
|
|
Ref<MLPPVector> MLPPLinReg::model_set_test(const Ref<MLPPMatrix> &X) {
|
|
ERR_FAIL_COND_V(!_initialized, Ref<MLPPVector>());
|
|
|
|
return evaluatem(X);
|
|
}
|
|
|
|
real_t MLPPLinReg::model_test(const Ref<MLPPVector> &x) {
|
|
ERR_FAIL_COND_V(!_initialized, 0);
|
|
|
|
return evaluatev(x);
|
|
}
|
|
|
|
void MLPPLinReg::newton_raphson(real_t learning_rate, int max_epoch, bool ui) {
|
|
ERR_FAIL_COND(!_initialized);
|
|
|
|
MLPPLinAlg alg;
|
|
MLPPReg regularization;
|
|
|
|
real_t cost_prev = 0;
|
|
int epoch = 1;
|
|
|
|
forward_pass();
|
|
|
|
while (true) {
|
|
cost_prev = cost(_y_hat, _output_set);
|
|
|
|
Ref<MLPPVector> error = alg.subtractionnv(_y_hat, _output_set);
|
|
|
|
// Calculating the weight gradients (2nd derivative)
|
|
Ref<MLPPVector> first_derivative = alg.mat_vec_multnv(alg.transposenm(_input_set), error);
|
|
Ref<MLPPMatrix> second_derivative = alg.matmultnm(alg.transposenm(_input_set), _input_set);
|
|
_weights = alg.subtractionnv(_weights, alg.scalar_multiplynv(learning_rate / _n, alg.mat_vec_multnv(alg.transposenm(alg.inversenm(second_derivative)), first_derivative)));
|
|
_weights = regularization.reg_weightsv(_weights, _lambda, _alpha, _reg);
|
|
|
|
// Calculating the bias gradients (2nd derivative)
|
|
_bias -= learning_rate * alg.sum_elementsv(error) / _n; // We keep this the same. The 2nd derivative is just [1].
|
|
|
|
forward_pass();
|
|
|
|
if (ui) {
|
|
MLPPUtilities::cost_info(epoch, cost_prev, cost(_y_hat, _output_set));
|
|
MLPPUtilities::print_ui_vb(_weights, _bias);
|
|
}
|
|
|
|
epoch++;
|
|
|
|
if (epoch > max_epoch) {
|
|
break;
|
|
}
|
|
}
|
|
}
|
|
|
|
void MLPPLinReg::gradient_descent(real_t learning_rate, int max_epoch, bool ui) {
|
|
ERR_FAIL_COND(!_initialized);
|
|
|
|
MLPPLinAlg alg;
|
|
MLPPReg regularization;
|
|
|
|
real_t cost_prev = 0;
|
|
int epoch = 1;
|
|
|
|
forward_pass();
|
|
|
|
while (true) {
|
|
cost_prev = cost(_y_hat, _output_set);
|
|
|
|
Ref<MLPPVector> error = alg.subtractionnv(_y_hat, _output_set);
|
|
|
|
// Calculating the weight gradients
|
|
_weights = alg.subtractionnv(_weights, alg.scalar_multiplynv(learning_rate / _n, alg.mat_vec_multnv(alg.transposenm(_input_set), error)));
|
|
_weights = regularization.reg_weightsv(_weights, _lambda, _alpha, _reg);
|
|
|
|
// Calculating the bias gradients
|
|
_bias -= learning_rate * alg.sum_elementsv(error) / _n;
|
|
|
|
forward_pass();
|
|
|
|
if (ui) {
|
|
MLPPUtilities::cost_info(epoch, cost_prev, cost(_y_hat, _output_set));
|
|
MLPPUtilities::print_ui_vb(_weights, _bias);
|
|
}
|
|
|
|
epoch++;
|
|
|
|
if (epoch > max_epoch) {
|
|
break;
|
|
}
|
|
}
|
|
}
|
|
|
|
void MLPPLinReg::sgd(real_t learning_rate, int max_epoch, bool ui) {
|
|
ERR_FAIL_COND(!_initialized);
|
|
|
|
MLPPLinAlg alg;
|
|
MLPPReg regularization;
|
|
|
|
real_t cost_prev = 0;
|
|
int epoch = 1;
|
|
|
|
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(1);
|
|
|
|
Ref<MLPPVector> y_hat_tmp;
|
|
y_hat_tmp.instance();
|
|
y_hat_tmp->resize(1);
|
|
|
|
while (true) {
|
|
int output_index = distribution(generator);
|
|
|
|
_input_set->get_row_into_mlpp_vector(output_index, input_set_row_tmp);
|
|
real_t output_element_set = _output_set->element_get(output_index);
|
|
output_set_row_tmp->element_set(0, output_element_set);
|
|
|
|
real_t y_hat = evaluatev(input_set_row_tmp);
|
|
y_hat_tmp->element_set(0, output_element_set);
|
|
|
|
cost_prev = cost(y_hat_tmp, output_set_row_tmp);
|
|
|
|
real_t error = y_hat - output_element_set;
|
|
|
|
// Weight updation
|
|
_weights = alg.subtractionnv(_weights, alg.scalar_multiplynv(learning_rate * error, input_set_row_tmp));
|
|
_weights = regularization.reg_weightsv(_weights, _lambda, _alpha, _reg);
|
|
|
|
// Bias updation
|
|
_bias -= learning_rate * error;
|
|
|
|
y_hat = evaluatev(input_set_row_tmp);
|
|
|
|
if (ui) {
|
|
MLPPUtilities::cost_info(epoch, cost_prev, cost(y_hat_tmp, output_set_row_tmp));
|
|
MLPPUtilities::print_ui_vb(_weights, _bias);
|
|
}
|
|
|
|
epoch++;
|
|
|
|
if (epoch > max_epoch) {
|
|
break;
|
|
}
|
|
}
|
|
|
|
forward_pass();
|
|
}
|
|
|
|
void MLPPLinReg::mbgd(real_t learning_rate, int max_epoch, int mini_batch_size, bool ui) {
|
|
ERR_FAIL_COND(!_initialized);
|
|
|
|
MLPPLinAlg alg;
|
|
MLPPReg regularization;
|
|
|
|
real_t cost_prev = 0;
|
|
int epoch = 1;
|
|
|
|
// Creating the mini-batches
|
|
int n_mini_batch = _n / mini_batch_size;
|
|
MLPPUtilities::CreateMiniBatchMVBatch batches = MLPPUtilities::create_mini_batchesmv(_input_set, _output_set, n_mini_batch);
|
|
|
|
while (true) {
|
|
for (int i = 0; i < n_mini_batch; i++) {
|
|
Ref<MLPPMatrix> current_input_mini_batch = batches.input_sets[i];
|
|
Ref<MLPPVector> current_output_mini_batch = batches.output_sets[i];
|
|
|
|
Ref<MLPPVector> y_hat = evaluatem(current_input_mini_batch);
|
|
cost_prev = cost(y_hat, current_output_mini_batch);
|
|
|
|
Ref<MLPPVector> error = alg.subtractionnv(y_hat, current_output_mini_batch);
|
|
|
|
// Calculating the weight gradients
|
|
_weights = alg.subtractionnv(_weights, alg.scalar_multiplynv(learning_rate / current_output_mini_batch->size(), alg.mat_vec_multnv(alg.transposenm(current_input_mini_batch), error)));
|
|
_weights = regularization.reg_weightsv(_weights, _lambda, _alpha, _reg);
|
|
|
|
// Calculating the bias gradients
|
|
_bias -= learning_rate * alg.sum_elementsv(error) / current_output_mini_batch->size();
|
|
y_hat = evaluatem(current_input_mini_batch);
|
|
|
|
if (ui) {
|
|
MLPPUtilities::cost_info(epoch, cost_prev, cost(y_hat, current_output_mini_batch));
|
|
MLPPUtilities::print_ui_vb(_weights, _bias);
|
|
}
|
|
}
|
|
|
|
epoch++;
|
|
|
|
if (epoch > max_epoch) {
|
|
break;
|
|
}
|
|
}
|
|
|
|
forward_pass();
|
|
}
|
|
|
|
void MLPPLinReg::momentum(real_t learning_rate, int max_epoch, int mini_batch_size, real_t gamma, bool ui) {
|
|
ERR_FAIL_COND(!_initialized);
|
|
|
|
MLPPLinAlg alg;
|
|
MLPPReg regularization;
|
|
real_t cost_prev = 0;
|
|
int epoch = 1;
|
|
|
|
// Creating the mini-batches
|
|
int n_mini_batch = _n / mini_batch_size;
|
|
MLPPUtilities::CreateMiniBatchMVBatch batches = MLPPUtilities::create_mini_batchesmv(_input_set, _output_set, n_mini_batch);
|
|
|
|
// Initializing necessary components for Momentum.
|
|
Ref<MLPPVector> v = alg.zerovecnv(_weights->size());
|
|
|
|
while (true) {
|
|
for (int i = 0; i < n_mini_batch; i++) {
|
|
Ref<MLPPMatrix> current_input_mini_batch = batches.input_sets[i];
|
|
Ref<MLPPVector> current_output_mini_batch = batches.output_sets[i];
|
|
|
|
Ref<MLPPVector> y_hat = evaluatem(current_input_mini_batch);
|
|
cost_prev = cost(y_hat, current_output_mini_batch);
|
|
|
|
Ref<MLPPVector> error = alg.subtractionnv(y_hat, current_output_mini_batch);
|
|
|
|
// Calculating the weight gradients
|
|
Ref<MLPPVector> gradient = alg.scalar_multiplynv(1 / current_output_mini_batch->size(), alg.mat_vec_multnv(alg.transposenm(current_input_mini_batch), error));
|
|
Ref<MLPPVector> reg_deriv_term = regularization.reg_deriv_termv(_weights, _lambda, _alpha, _reg);
|
|
Ref<MLPPVector> weight_grad = alg.additionnv(gradient, reg_deriv_term); // Weight_grad_final
|
|
|
|
v = alg.additionnv(alg.scalar_multiplynv(gamma, v), alg.scalar_multiplynv(learning_rate, weight_grad));
|
|
|
|
_weights = alg.subtractionnv(_weights, v);
|
|
|
|
// Calculating the bias gradients
|
|
_bias -= learning_rate * alg.sum_elementsv(error) / current_output_mini_batch->size(); // As normal
|
|
y_hat = evaluatem(current_input_mini_batch);
|
|
|
|
if (ui) {
|
|
MLPPUtilities::cost_info(epoch, cost_prev, cost(y_hat, current_output_mini_batch));
|
|
MLPPUtilities::print_ui_vb(_weights, _bias);
|
|
}
|
|
}
|
|
|
|
epoch++;
|
|
|
|
if (epoch > max_epoch) {
|
|
break;
|
|
}
|
|
}
|
|
|
|
forward_pass();
|
|
}
|
|
|
|
void MLPPLinReg::nag(real_t learning_rate, int max_epoch, int mini_batch_size, real_t gamma, bool ui) {
|
|
ERR_FAIL_COND(!_initialized);
|
|
|
|
MLPPLinAlg alg;
|
|
MLPPReg regularization;
|
|
real_t cost_prev = 0;
|
|
int epoch = 1;
|
|
|
|
// Creating the mini-batches
|
|
int n_mini_batch = _n / mini_batch_size;
|
|
MLPPUtilities::CreateMiniBatchMVBatch batches = MLPPUtilities::create_mini_batchesmv(_input_set, _output_set, n_mini_batch);
|
|
|
|
// Initializing necessary components for Momentum.
|
|
Ref<MLPPVector> v = alg.zerovecnv(_weights->size());
|
|
|
|
while (true) {
|
|
for (int i = 0; i < n_mini_batch; i++) {
|
|
Ref<MLPPMatrix> current_input_mini_batch = batches.input_sets[i];
|
|
Ref<MLPPVector> current_output_mini_batch = batches.output_sets[i];
|
|
|
|
_weights = alg.subtractionnv(_weights, alg.scalar_multiplynv(gamma, v)); // "Aposterori" calculation
|
|
|
|
Ref<MLPPVector> y_hat = evaluatem(current_input_mini_batch);
|
|
cost_prev = cost(y_hat, current_output_mini_batch);
|
|
|
|
Ref<MLPPVector> error = alg.subtractionnv(y_hat, current_output_mini_batch);
|
|
|
|
// Calculating the weight gradients
|
|
Ref<MLPPVector> gradient = alg.scalar_multiplynv(1 / current_output_mini_batch->size(), alg.mat_vec_multnv(alg.transposenm(current_input_mini_batch), error));
|
|
Ref<MLPPVector> reg_deriv_term = regularization.reg_deriv_termv(_weights, _lambda, _alpha, _reg);
|
|
Ref<MLPPVector> weight_grad = alg.additionnv(gradient, reg_deriv_term); // Weight_grad_final
|
|
|
|
v = alg.additionnv(alg.scalar_multiplynv(gamma, v), alg.scalar_multiplynv(learning_rate, weight_grad));
|
|
|
|
_weights = alg.subtractionnv(_weights, v);
|
|
|
|
// Calculating the bias gradients
|
|
_bias -= learning_rate * alg.sum_elementsv(error) / current_output_mini_batch->size(); // As normal
|
|
y_hat = evaluatem(current_input_mini_batch);
|
|
|
|
if (ui) {
|
|
MLPPUtilities::cost_info(epoch, cost_prev, cost(y_hat, current_output_mini_batch));
|
|
MLPPUtilities::print_ui_vb(_weights, _bias);
|
|
}
|
|
}
|
|
|
|
epoch++;
|
|
|
|
if (epoch > max_epoch) {
|
|
break;
|
|
}
|
|
}
|
|
|
|
forward_pass();
|
|
}
|
|
|
|
void MLPPLinReg::adagrad(real_t learning_rate, int max_epoch, int mini_batch_size, real_t e, bool ui) {
|
|
ERR_FAIL_COND(!_initialized);
|
|
|
|
MLPPLinAlg alg;
|
|
MLPPReg regularization;
|
|
real_t cost_prev = 0;
|
|
int epoch = 1;
|
|
|
|
// Creating the mini-batches
|
|
int n_mini_batch = _n / mini_batch_size;
|
|
MLPPUtilities::CreateMiniBatchMVBatch batches = MLPPUtilities::create_mini_batchesmv(_input_set, _output_set, n_mini_batch);
|
|
|
|
// Initializing necessary components for Adagrad.
|
|
Ref<MLPPVector> v = alg.zerovecnv(_weights->size());
|
|
|
|
while (true) {
|
|
for (int i = 0; i < n_mini_batch; i++) {
|
|
Ref<MLPPMatrix> current_input_mini_batch = batches.input_sets[i];
|
|
Ref<MLPPVector> current_output_mini_batch = batches.output_sets[i];
|
|
|
|
Ref<MLPPVector> y_hat = evaluatem(current_input_mini_batch);
|
|
cost_prev = cost(y_hat, current_output_mini_batch);
|
|
|
|
Ref<MLPPVector> error = alg.subtractionnv(y_hat, current_output_mini_batch);
|
|
|
|
// Calculating the weight gradients
|
|
Ref<MLPPVector> gradient = alg.scalar_multiplynv(1 / current_output_mini_batch->size(), alg.mat_vec_multnv(alg.transposenm(current_input_mini_batch), error));
|
|
Ref<MLPPVector> reg_deriv_term = regularization.reg_deriv_termv(_weights, _lambda, _alpha, _reg);
|
|
Ref<MLPPVector> weight_grad = alg.additionnv(gradient, reg_deriv_term); // Weight_grad_final
|
|
|
|
v = alg.hadamard_productnv(weight_grad, weight_grad);
|
|
|
|
_weights = alg.subtractionnv(_weights, alg.scalar_multiplynv(learning_rate, alg.division_element_wisenv(weight_grad, alg.sqrtnv(alg.scalar_addnv(e, v)))));
|
|
|
|
// Calculating the bias gradients
|
|
_bias -= learning_rate * alg.sum_elementsv(error) / current_output_mini_batch->size(); // As normal
|
|
y_hat = evaluatem(current_input_mini_batch);
|
|
|
|
if (ui) {
|
|
MLPPUtilities::cost_info(epoch, cost_prev, cost(y_hat, current_output_mini_batch));
|
|
MLPPUtilities::print_ui_vb(_weights, _bias);
|
|
}
|
|
}
|
|
|
|
epoch++;
|
|
|
|
if (epoch > max_epoch) {
|
|
break;
|
|
}
|
|
}
|
|
|
|
forward_pass();
|
|
}
|
|
|
|
void MLPPLinReg::adadelta(real_t learning_rate, int max_epoch, int mini_batch_size, real_t b1, real_t e, bool ui) {
|
|
ERR_FAIL_COND(!_initialized);
|
|
|
|
// Adagrad upgrade. Momentum is applied.
|
|
MLPPLinAlg alg;
|
|
MLPPReg regularization;
|
|
real_t cost_prev = 0;
|
|
int epoch = 1;
|
|
|
|
// Creating the mini-batches
|
|
int n_mini_batch = _n / mini_batch_size;
|
|
MLPPUtilities::CreateMiniBatchMVBatch batches = MLPPUtilities::create_mini_batchesmv(_input_set, _output_set, n_mini_batch);
|
|
|
|
// Initializing necessary components for Adagrad.
|
|
Ref<MLPPVector> v = alg.zerovecnv(_weights->size());
|
|
|
|
while (true) {
|
|
for (int i = 0; i < n_mini_batch; i++) {
|
|
Ref<MLPPMatrix> current_input_mini_batch = batches.input_sets[i];
|
|
Ref<MLPPVector> current_output_mini_batch = batches.output_sets[i];
|
|
|
|
Ref<MLPPVector> y_hat = evaluatem(current_input_mini_batch);
|
|
cost_prev = cost(y_hat, current_output_mini_batch);
|
|
|
|
Ref<MLPPVector> error = alg.subtractionnv(y_hat, current_output_mini_batch);
|
|
|
|
// Calculating the weight gradients
|
|
Ref<MLPPVector> gradient = alg.scalar_multiplynv(1 / current_output_mini_batch->size(), alg.mat_vec_multnv(alg.transposenm(current_input_mini_batch), error));
|
|
Ref<MLPPVector> reg_deriv_term = regularization.reg_deriv_termv(_weights, _lambda, _alpha, _reg);
|
|
Ref<MLPPVector> weight_grad = alg.additionnv(gradient, reg_deriv_term); // Weight_grad_final
|
|
|
|
v = alg.additionnv(alg.scalar_multiplynv(b1, v), alg.scalar_multiplynv(1 - b1, alg.hadamard_productnv(weight_grad, weight_grad)));
|
|
|
|
_weights = alg.subtractionnv(_weights, alg.scalar_multiplynv(learning_rate, alg.division_element_wisenv(weight_grad, alg.sqrtnv(alg.scalar_addnv(e, v)))));
|
|
|
|
// Calculating the bias gradients
|
|
_bias -= learning_rate * alg.sum_elementsv(error) / current_output_mini_batch->size(); // As normal
|
|
y_hat = evaluatem(current_input_mini_batch);
|
|
|
|
if (ui) {
|
|
MLPPUtilities::cost_info(epoch, cost_prev, cost(y_hat, current_output_mini_batch));
|
|
MLPPUtilities::print_ui_vb(_weights, _bias);
|
|
}
|
|
}
|
|
|
|
epoch++;
|
|
|
|
if (epoch > max_epoch) {
|
|
break;
|
|
}
|
|
}
|
|
|
|
forward_pass();
|
|
}
|
|
|
|
void MLPPLinReg::adam(real_t learning_rate, int max_epoch, int mini_batch_size, real_t b1, real_t b2, real_t e, bool ui) {
|
|
ERR_FAIL_COND(!_initialized);
|
|
|
|
MLPPLinAlg alg;
|
|
MLPPReg regularization;
|
|
real_t cost_prev = 0;
|
|
int epoch = 1;
|
|
|
|
// Creating the mini-batches
|
|
int n_mini_batch = _n / mini_batch_size;
|
|
MLPPUtilities::CreateMiniBatchMVBatch batches = MLPPUtilities::create_mini_batchesmv(_input_set, _output_set, n_mini_batch);
|
|
|
|
// Initializing necessary components for Adam.
|
|
Ref<MLPPVector> m = alg.zerovecnv(_weights->size());
|
|
Ref<MLPPVector> v = alg.zerovecnv(_weights->size());
|
|
|
|
while (true) {
|
|
for (int i = 0; i < n_mini_batch; i++) {
|
|
Ref<MLPPMatrix> current_input_mini_batch = batches.input_sets[i];
|
|
Ref<MLPPVector> current_output_mini_batch = batches.output_sets[i];
|
|
|
|
Ref<MLPPVector> y_hat = evaluatem(current_input_mini_batch);
|
|
cost_prev = cost(y_hat, current_output_mini_batch);
|
|
|
|
Ref<MLPPVector> error = alg.subtractionnv(y_hat, current_output_mini_batch);
|
|
|
|
// Calculating the weight gradients
|
|
Ref<MLPPVector> gradient = alg.scalar_multiplynv(1 / current_output_mini_batch->size(), alg.mat_vec_multnv(alg.transposenm(current_input_mini_batch), error));
|
|
Ref<MLPPVector> reg_deriv_term = regularization.reg_deriv_termv(_weights, _lambda, _alpha, _reg);
|
|
Ref<MLPPVector> weight_grad = alg.additionnv(gradient, reg_deriv_term); // Weight_grad_final
|
|
|
|
m = alg.additionnv(alg.scalar_multiplynv(b1, m), alg.scalar_multiplynv(1 - b1, weight_grad));
|
|
v = alg.additionnv(alg.scalar_multiplynv(b2, v), alg.scalar_multiplynv(1 - b2, alg.exponentiatenv(weight_grad, 2)));
|
|
|
|
Ref<MLPPVector> m_hat = alg.scalar_multiplynv(1 / (1 - Math::pow(b1, epoch)), m);
|
|
Ref<MLPPVector> v_hat = alg.scalar_multiplynv(1 / (1 - Math::pow(b2, epoch)), v);
|
|
|
|
_weights = alg.subtractionnv(_weights, alg.scalar_multiplynv(learning_rate, alg.division_element_wisenvnm(m_hat, alg.scalar_addnv(e, alg.sqrtnv(v_hat)))));
|
|
|
|
// Calculating the bias gradients
|
|
_bias -= learning_rate * alg.sum_elementsv(error) / current_output_mini_batch->size(); // As normal
|
|
y_hat = evaluatem(current_input_mini_batch);
|
|
|
|
if (ui) {
|
|
MLPPUtilities::cost_info(epoch, cost_prev, cost(y_hat, current_output_mini_batch));
|
|
MLPPUtilities::print_ui_vb(_weights, _bias);
|
|
}
|
|
}
|
|
|
|
epoch++;
|
|
|
|
if (epoch > max_epoch) {
|
|
break;
|
|
}
|
|
}
|
|
|
|
forward_pass();
|
|
}
|
|
|
|
void MLPPLinReg::adamax(real_t learning_rate, int max_epoch, int mini_batch_size, real_t b1, real_t b2, real_t e, bool ui) {
|
|
ERR_FAIL_COND(!_initialized);
|
|
|
|
MLPPLinAlg alg;
|
|
MLPPReg regularization;
|
|
real_t cost_prev = 0;
|
|
int epoch = 1;
|
|
|
|
// Creating the mini-batches
|
|
int n_mini_batch = _n / mini_batch_size;
|
|
MLPPUtilities::CreateMiniBatchMVBatch batches = MLPPUtilities::create_mini_batchesmv(_input_set, _output_set, n_mini_batch);
|
|
|
|
Ref<MLPPVector> m = alg.zerovecnv(_weights->size());
|
|
Ref<MLPPVector> u = alg.zerovecnv(_weights->size());
|
|
|
|
while (true) {
|
|
for (int i = 0; i < n_mini_batch; i++) {
|
|
Ref<MLPPMatrix> current_input_mini_batch = batches.input_sets[i];
|
|
Ref<MLPPVector> current_output_mini_batch = batches.output_sets[i];
|
|
|
|
Ref<MLPPVector> y_hat = evaluatem(current_input_mini_batch);
|
|
cost_prev = cost(y_hat, current_output_mini_batch);
|
|
|
|
Ref<MLPPVector> error = alg.subtractionnv(y_hat, current_output_mini_batch);
|
|
|
|
// Calculating the weight gradients
|
|
Ref<MLPPVector> gradient = alg.scalar_multiplynv(1 / current_output_mini_batch->size(), alg.mat_vec_multnv(alg.transposenm(current_input_mini_batch), error));
|
|
Ref<MLPPVector> reg_deriv_term = regularization.reg_deriv_termv(_weights, _lambda, _alpha, _reg);
|
|
Ref<MLPPVector> weight_grad = alg.additionnv(gradient, reg_deriv_term); // Weight_grad_final
|
|
|
|
m = alg.additionnv(alg.scalar_multiplynv(b1, m), alg.scalar_multiplynv(1 - b1, weight_grad));
|
|
u = alg.maxnvv(alg.scalar_multiplynv(b2, u), alg.absv(weight_grad));
|
|
|
|
Ref<MLPPVector> m_hat = alg.scalar_multiplynv(1 / (1 - Math::pow(b1, epoch)), m);
|
|
|
|
_weights = alg.subtractionnv(_weights, alg.scalar_multiplynv(learning_rate, alg.division_element_wisenv(m_hat, u)));
|
|
|
|
// Calculating the bias gradients
|
|
_bias -= learning_rate * alg.sum_elementsv(error) / current_output_mini_batch->size(); // As normal
|
|
y_hat = evaluatem(current_input_mini_batch);
|
|
|
|
if (ui) {
|
|
MLPPUtilities::cost_info(epoch, cost_prev, cost(y_hat, current_output_mini_batch));
|
|
MLPPUtilities::print_ui_vb(_weights, _bias);
|
|
}
|
|
}
|
|
|
|
epoch++;
|
|
|
|
if (epoch > max_epoch) {
|
|
break;
|
|
}
|
|
}
|
|
|
|
forward_pass();
|
|
}
|
|
|
|
void MLPPLinReg::nadam(real_t learning_rate, int max_epoch, int mini_batch_size, real_t b1, real_t b2, real_t e, bool ui) {
|
|
ERR_FAIL_COND(!_initialized);
|
|
|
|
MLPPLinAlg alg;
|
|
MLPPReg regularization;
|
|
real_t cost_prev = 0;
|
|
int epoch = 1;
|
|
|
|
// Creating the mini-batches
|
|
int n_mini_batch = _n / mini_batch_size;
|
|
MLPPUtilities::CreateMiniBatchMVBatch batches = MLPPUtilities::create_mini_batchesmv(_input_set, _output_set, n_mini_batch);
|
|
|
|
// Initializing necessary components for Adam.
|
|
Ref<MLPPVector> m = alg.zerovecnv(_weights->size());
|
|
Ref<MLPPVector> v = alg.zerovecnv(_weights->size());
|
|
Ref<MLPPVector> m_final = alg.zerovecnv(_weights->size());
|
|
|
|
while (true) {
|
|
for (int i = 0; i < n_mini_batch; i++) {
|
|
Ref<MLPPMatrix> current_input_mini_batch = batches.input_sets[i];
|
|
Ref<MLPPVector> current_output_mini_batch = batches.output_sets[i];
|
|
|
|
Ref<MLPPVector> y_hat = evaluatem(current_input_mini_batch);
|
|
cost_prev = cost(y_hat, current_output_mini_batch);
|
|
|
|
Ref<MLPPVector> error = alg.subtractionnv(y_hat, current_output_mini_batch);
|
|
|
|
// Calculating the weight gradients
|
|
Ref<MLPPVector> gradient = alg.scalar_multiplynv(1 / current_output_mini_batch->size(), alg.mat_vec_multnv(alg.transposenm(current_input_mini_batch), error));
|
|
Ref<MLPPVector> reg_deriv_term = regularization.reg_deriv_termv(_weights, _lambda, _alpha, _reg);
|
|
Ref<MLPPVector> weight_grad = alg.additionnv(gradient, reg_deriv_term); // Weight_grad_final
|
|
|
|
m = alg.additionnv(alg.scalar_multiplynv(b1, m), alg.scalar_multiplynv(1 - b1, weight_grad));
|
|
v = alg.additionnv(alg.scalar_multiplynv(b2, v), alg.scalar_multiplynv(1 - b2, alg.exponentiatenv(weight_grad, 2)));
|
|
m_final = alg.additionnv(alg.scalar_multiplynv(b1, m), alg.scalar_multiplynv((1 - b1) / (1 - Math::pow(b1, epoch)), weight_grad));
|
|
|
|
Ref<MLPPVector> m_hat = alg.scalar_multiplynv(1 / (1 - Math::pow(b1, epoch)), m);
|
|
Ref<MLPPVector> v_hat = alg.scalar_multiplynv(1 / (1 - Math::pow(b2, epoch)), v);
|
|
|
|
_weights = alg.subtractionnv(_weights, alg.scalar_multiplynv(learning_rate, alg.division_element_wisenv(m_final, alg.scalar_addnv(e, alg.sqrtnv(v_hat)))));
|
|
|
|
// Calculating the bias gradients
|
|
_bias -= learning_rate * alg.sum_elementsv(error) / current_output_mini_batch->size(); // As normal
|
|
y_hat = evaluatem(current_input_mini_batch);
|
|
|
|
if (ui) {
|
|
MLPPUtilities::cost_info(epoch, cost_prev, cost(y_hat, current_output_mini_batch));
|
|
MLPPUtilities::print_ui_vb(_weights, _bias);
|
|
}
|
|
}
|
|
|
|
epoch++;
|
|
|
|
if (epoch > max_epoch) {
|
|
break;
|
|
}
|
|
}
|
|
|
|
forward_pass();
|
|
}
|
|
|
|
void MLPPLinReg::normal_equation() {
|
|
ERR_FAIL_COND(!_initialized);
|
|
|
|
MLPPLinAlg alg;
|
|
MLPPStat stat;
|
|
|
|
Ref<MLPPMatrix> input_set_t = alg.transposenm(_input_set);
|
|
|
|
Ref<MLPPVector> input_set_t_row_tmp;
|
|
input_set_t_row_tmp.instance();
|
|
input_set_t_row_tmp->resize(input_set_t->size().x);
|
|
|
|
Ref<MLPPVector> x_means;
|
|
x_means.instance();
|
|
x_means->resize(input_set_t->size().y);
|
|
|
|
for (int i = 0; i < input_set_t->size().y; i++) {
|
|
input_set_t->get_row_into_mlpp_vector(i, input_set_t_row_tmp);
|
|
|
|
x_means->element_set(i, stat.meanv(input_set_t_row_tmp));
|
|
}
|
|
|
|
Ref<MLPPVector> temp;
|
|
//temp.resize(_k);
|
|
temp = alg.mat_vec_multnv(alg.inversenm(alg.matmultnm(alg.transposenm(_input_set), _input_set)), alg.mat_vec_multnv(alg.transposenm(_input_set), _output_set));
|
|
|
|
ERR_FAIL_COND_MSG(Math::is_nan(temp->element_get(0)), "ERR: Resulting matrix was noninvertible/degenerate, and so the normal equation could not be performed. Try utilizing gradient descent.");
|
|
|
|
if (_reg == MLPPReg::REGULARIZATION_TYPE_RIDGE) {
|
|
_weights = alg.mat_vec_multnv(alg.inversenm(alg.additionnm(alg.matmultnm(alg.transposenm(_input_set), _input_set), alg.scalar_multiplynm(_lambda, alg.identitym(_k)))), alg.mat_vec_multnv(alg.transposenm(_input_set), _output_set));
|
|
} else {
|
|
_weights = alg.mat_vec_multnv(alg.inversenm(alg.matmultnm(alg.transposenm(_input_set), _input_set)), alg.mat_vec_multnv(alg.transposenm(_input_set), _output_set));
|
|
}
|
|
|
|
_bias = stat.meanv(_output_set) - alg.dotnv(_weights, x_means);
|
|
|
|
forward_pass();
|
|
}
|
|
|
|
real_t MLPPLinReg::score() {
|
|
ERR_FAIL_COND_V(!_initialized, 0);
|
|
|
|
MLPPUtilities util;
|
|
|
|
return util.performance_vec(_y_hat, _output_set);
|
|
}
|
|
|
|
void MLPPLinReg::save(const String &file_name) {
|
|
ERR_FAIL_COND(!_initialized);
|
|
|
|
//MLPPUtilities util;
|
|
|
|
//util.saveParameters(fileName, _weights, _bias);
|
|
}
|
|
|
|
bool MLPPLinReg::is_initialized() {
|
|
return _initialized;
|
|
}
|
|
void MLPPLinReg::initialize() {
|
|
if (_initialized) {
|
|
return;
|
|
}
|
|
|
|
//ERR_FAIL_COND(!_input_set.is_valid() || !_output_set.is_valid());
|
|
|
|
_initialized = true;
|
|
}
|
|
|
|
MLPPLinReg::MLPPLinReg(const Ref<MLPPMatrix> &p_input_set, const Ref<MLPPVector> &p_output_set, 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().y;
|
|
_k = p_input_set->size().x;
|
|
_reg = p_reg;
|
|
_lambda = p_lambda;
|
|
_alpha = p_alpha;
|
|
|
|
_y_hat.instance();
|
|
_y_hat->resize(_n);
|
|
|
|
_weights.instance();
|
|
_weights->resize(_k);
|
|
|
|
MLPPUtilities utils;
|
|
|
|
utils.weight_initializationv(_weights);
|
|
_bias = utils.bias_initializationr();
|
|
|
|
_initialized = true;
|
|
}
|
|
|
|
MLPPLinReg::MLPPLinReg() {
|
|
_initialized = false;
|
|
}
|
|
MLPPLinReg::~MLPPLinReg() {
|
|
}
|
|
|
|
real_t MLPPLinReg::cost(const Ref<MLPPVector> &y_hat, const Ref<MLPPVector> &y) {
|
|
MLPPReg regularization;
|
|
MLPPCost mlpp_cost;
|
|
|
|
return mlpp_cost.msev(y_hat, y) + regularization.reg_termv(_weights, _lambda, _alpha, _reg);
|
|
}
|
|
|
|
real_t MLPPLinReg::evaluatev(const Ref<MLPPVector> &x) {
|
|
MLPPLinAlg alg;
|
|
|
|
return alg.dotnv(_weights, x) + _bias;
|
|
}
|
|
|
|
Ref<MLPPVector> MLPPLinReg::evaluatem(const Ref<MLPPMatrix> &X) {
|
|
MLPPLinAlg alg;
|
|
|
|
return alg.scalar_addnv(_bias, alg.mat_vec_multnv(X, _weights));
|
|
}
|
|
|
|
// wTx + b
|
|
void MLPPLinReg::forward_pass() {
|
|
_y_hat = evaluatem(_input_set);
|
|
}
|
|
|
|
void MLPPLinReg::_bind_methods() {
|
|
/*
|
|
ClassDB::bind_method(D_METHOD("get_input_set"), &MLPPLinReg::get_input_set);
|
|
ClassDB::bind_method(D_METHOD("set_input_set", "val"), &MLPPLinReg::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"), &MLPPLinReg::get_output_set);
|
|
ClassDB::bind_method(D_METHOD("set_output_set", "val"), &MLPPLinReg::set_output_set);
|
|
ADD_PROPERTY(PropertyInfo(Variant::OBJECT, "output_set", PROPERTY_HINT_RESOURCE_TYPE, "MLPPVector"), "set_output_set", "get_output_set");
|
|
|
|
ClassDB::bind_method(D_METHOD("get_reg"), &MLPPLinReg::get_reg);
|
|
ClassDB::bind_method(D_METHOD("set_reg", "val"), &MLPPLinReg::set_reg);
|
|
ADD_PROPERTY(PropertyInfo(Variant::INT, "reg"), "set_reg", "get_reg");
|
|
|
|
ClassDB::bind_method(D_METHOD("get_lambda"), &MLPPLinReg::get_lambda);
|
|
ClassDB::bind_method(D_METHOD("set_lambda", "val"), &MLPPLinReg::set_lambda);
|
|
ADD_PROPERTY(PropertyInfo(Variant::REAL, "lambda"), "set_lambda", "get_lambda");
|
|
|
|
ClassDB::bind_method(D_METHOD("get_alpha"), &MLPPLinReg::get_alpha);
|
|
ClassDB::bind_method(D_METHOD("set_alpha", "val"), &MLPPLinReg::set_alpha);
|
|
ADD_PROPERTY(PropertyInfo(Variant::REAL, "alpha"), "set_alpha", "get_alpha");
|
|
|
|
ClassDB::bind_method(D_METHOD("model_test", "x"), &MLPPLinReg::model_test);
|
|
ClassDB::bind_method(D_METHOD("model_set_test", "X"), &MLPPLinReg::model_set_test);
|
|
|
|
ClassDB::bind_method(D_METHOD("gradient_descent", "learning_rate", "max_epoch", "ui"), &MLPPLinReg::gradient_descent, false);
|
|
ClassDB::bind_method(D_METHOD("sgd", "learning_rate", "max_epoch", "ui"), &MLPPLinReg::sgd, false);
|
|
ClassDB::bind_method(D_METHOD("mbgd", "learning_rate", "max_epoch", "mini_batch_size", "ui"), &MLPPLinReg::mbgd, false);
|
|
|
|
ClassDB::bind_method(D_METHOD("score"), &MLPPLinReg::score);
|
|
|
|
ClassDB::bind_method(D_METHOD("save", "file_name"), &MLPPLinReg::save);
|
|
|
|
ClassDB::bind_method(D_METHOD("is_initialized"), &MLPPLinReg::is_initialized);
|
|
ClassDB::bind_method(D_METHOD("initialize"), &MLPPLinReg::initialize);
|
|
*/
|
|
}
|