pmlpp/mlpp/ann/ann.cpp
2023-02-13 00:56:09 +01:00

857 lines
32 KiB
C++

//
// ANN.cpp
//
// Created by Marc Melikyan on 11/4/20.
//
#include "ann.h"
#include "../activation/activation.h"
#include "../cost/cost.h"
#include "../lin_alg/lin_alg.h"
#include "../regularization/reg.h"
#include "../utilities/utilities.h"
#include <cmath>
#include <iostream>
#include <random>
std::vector<real_t> MLPPANN::model_set_test(std::vector<std::vector<real_t>> X) {
if (!_network.empty()) {
_network[0].input = X;
_network[0].forwardPass();
for (uint32_t i = 1; i < _network.size(); i++) {
_network[i].input = _network[i - 1].a;
_network[i].forwardPass();
}
_output_layer->input = _network[_network.size() - 1].a;
} else {
_output_layer->input = X;
}
_output_layer->forwardPass();
return _output_layer->a;
}
real_t MLPPANN::model_test(std::vector<real_t> x) {
if (!_network.empty()) {
_network[0].Test(x);
for (uint32_t i = 1; i < _network.size(); i++) {
_network[i].Test(_network[i - 1].a_test);
}
_output_layer->Test(_network[_network.size() - 1].a_test);
} else {
_output_layer->Test(x);
}
return _output_layer->a_test;
}
void MLPPANN::gradient_descent(real_t learning_rate, int max_epoch, bool ui) {
MLPPCost mlpp_cost;
MLPPLinAlg alg;
real_t cost_prev = 0;
int epoch = 1;
forward_pass();
real_t initial_learning_rate = learning_rate;
alg.printMatrix(_network[_network.size() - 1].weights);
while (true) {
learning_rate = apply_learning_rate_scheduler(initial_learning_rate, _decay_constant, epoch, _drop_rate);
cost_prev = cost(_y_hat, _output_set);
auto grads = compute_gradients(_y_hat, _output_set);
auto cumulative_hidden_layer_w_grad = std::get<0>(grads);
auto output_w_grad = std::get<1>(grads);
cumulative_hidden_layer_w_grad = alg.scalarMultiply(learning_rate / _n, cumulative_hidden_layer_w_grad);
output_w_grad = alg.scalarMultiply(learning_rate / _n, output_w_grad);
update_parameters(cumulative_hidden_layer_w_grad, output_w_grad, learning_rate); // subject to change. may want bias to have this matrix too.
std::cout << learning_rate << std::endl;
forward_pass();
if (ui) {
print_ui(epoch, cost_prev, _y_hat, _output_set);
}
epoch++;
if (epoch > max_epoch) {
break;
}
}
}
void MLPPANN::sgd(real_t learning_rate, int max_epoch, bool ui) {
MLPPCost mlpp_cost;
MLPPLinAlg alg;
real_t cost_prev = 0;
int epoch = 1;
real_t initial_learning_rate = learning_rate;
while (true) {
learning_rate = apply_learning_rate_scheduler(initial_learning_rate, _decay_constant, epoch, _drop_rate);
std::random_device rd;
std::default_random_engine generator(rd());
std::uniform_int_distribution<int> distribution(0, int(_n - 1));
int outputIndex = distribution(generator);
std::vector<real_t> y_hat = model_set_test({ _input_set[outputIndex] });
cost_prev = cost({ y_hat }, { _output_set[outputIndex] });
auto grads = compute_gradients(y_hat, { _output_set[outputIndex] });
auto cumulative_hidden_layer_w_grad = std::get<0>(grads);
auto output_w_grad = std::get<1>(grads);
cumulative_hidden_layer_w_grad = alg.scalarMultiply(learning_rate / _n, cumulative_hidden_layer_w_grad);
output_w_grad = alg.scalarMultiply(learning_rate / _n, output_w_grad);
update_parameters(cumulative_hidden_layer_w_grad, output_w_grad, learning_rate); // subject to change. may want bias to have this matrix too.
y_hat = model_set_test({ _input_set[outputIndex] });
if (ui) {
print_ui(epoch, cost_prev, y_hat, { _output_set[outputIndex] });
}
epoch++;
if (epoch > max_epoch) {
break;
}
}
forward_pass();
}
void MLPPANN::mbgd(real_t learning_rate, int max_epoch, int mini_batch_size, bool ui) {
MLPPCost mlpp_cost;
MLPPLinAlg alg;
real_t cost_prev = 0;
int epoch = 1;
real_t initial_learning_rate = learning_rate;
// Creating the mini-batches
int n_mini_batch = _n / mini_batch_size;
// always evaluate the result
// always do forward pass only ONCE at end.
auto batches = MLPPUtilities::createMiniBatches(_input_set, _output_set, n_mini_batch);
auto input_mini_batches = std::get<0>(batches);
auto output_mini_batches = std::get<1>(batches);
while (true) {
learning_rate = apply_learning_rate_scheduler(initial_learning_rate, _decay_constant, epoch, _drop_rate);
for (int i = 0; i < n_mini_batch; i++) {
std::vector<real_t> y_hat = model_set_test(input_mini_batches[i]);
cost_prev = cost(y_hat, output_mini_batches[i]);
auto grads = compute_gradients(y_hat, output_mini_batches[i]);
auto cumulative_hidden_layer_w_grad = std::get<0>(grads);
auto output_w_grad = std::get<1>(grads);
cumulative_hidden_layer_w_grad = alg.scalarMultiply(learning_rate / _n, cumulative_hidden_layer_w_grad);
output_w_grad = alg.scalarMultiply(learning_rate / _n, output_w_grad);
update_parameters(cumulative_hidden_layer_w_grad, output_w_grad, learning_rate); // subject to change. may want bias to have this matrix too.
y_hat = model_set_test(input_mini_batches[i]);
if (ui) {
print_ui(epoch, cost_prev, y_hat, output_mini_batches[i]);
}
}
epoch++;
if (epoch > max_epoch) {
break;
}
}
forward_pass();
}
void MLPPANN::momentum(real_t learning_rate, int max_epoch, int mini_batch_size, real_t gamma, bool nag, bool ui) {
class MLPPCost mlpp_cost;
MLPPLinAlg alg;
real_t cost_prev = 0;
int epoch = 1;
real_t initial_learning_rate = learning_rate;
// Creating the mini-batches
int n_mini_batch = _n / mini_batch_size;
// always evaluate the result
// always do forward pass only ONCE at end.
auto batches = MLPPUtilities::createMiniBatches(_input_set, _output_set, n_mini_batch);
auto input_mini_batches = std::get<0>(batches);
auto output_mini_batches = std::get<1>(batches);
// Initializing necessary components for Adam.
std::vector<std::vector<std::vector<real_t>>> v_hidden;
std::vector<real_t> v_output;
while (true) {
learning_rate = apply_learning_rate_scheduler(initial_learning_rate, _decay_constant, epoch, _drop_rate);
for (int i = 0; i < n_mini_batch; i++) {
std::vector<real_t> y_hat = model_set_test(input_mini_batches[i]);
cost_prev = cost(y_hat, output_mini_batches[i]);
auto grads = compute_gradients(y_hat, output_mini_batches[i]);
auto cumulative_hidden_layer_w_grad = std::get<0>(grads);
auto output_w_grad = std::get<1>(grads);
if (!_network.empty() && v_hidden.empty()) { // Initing our tensor
v_hidden = alg.resize(v_hidden, cumulative_hidden_layer_w_grad);
}
if (v_output.empty()) {
v_output.resize(output_w_grad.size());
}
if (nag) { // "Aposterori" calculation
update_parameters(v_hidden, v_output, 0); // DON'T update bias.
}
v_hidden = alg.addition(alg.scalarMultiply(gamma, v_hidden), alg.scalarMultiply(learning_rate / _n, cumulative_hidden_layer_w_grad));
v_output = alg.addition(alg.scalarMultiply(gamma, v_output), alg.scalarMultiply(learning_rate / _n, output_w_grad));
update_parameters(v_hidden, v_output, learning_rate); // subject to change. may want bias to have this matrix too.
y_hat = model_set_test(input_mini_batches[i]);
if (ui) {
print_ui(epoch, cost_prev, y_hat, output_mini_batches[i]);
}
}
epoch++;
if (epoch > max_epoch) {
break;
}
}
forward_pass();
}
void MLPPANN::adagrad(real_t learning_rate, int max_epoch, int mini_batch_size, real_t e, bool ui) {
MLPPCost mlpp_cost;
MLPPLinAlg alg;
real_t cost_prev = 0;
int epoch = 1;
real_t initial_learning_rate = learning_rate;
// Creating the mini-batches
int n_mini_batch = _n / mini_batch_size;
// always evaluate the result
// always do forward pass only ONCE at end.
auto batches = MLPPUtilities::createMiniBatches(_input_set, _output_set, n_mini_batch);
auto input_mini_batches = std::get<0>(batches);
auto output_mini_batches = std::get<1>(batches);
// Initializing necessary components for Adam.
std::vector<std::vector<std::vector<real_t>>> v_hidden;
std::vector<real_t> v_output;
while (true) {
learning_rate = apply_learning_rate_scheduler(initial_learning_rate, _decay_constant, epoch, _drop_rate);
for (int i = 0; i < n_mini_batch; i++) {
std::vector<real_t> y_hat = model_set_test(input_mini_batches[i]);
cost_prev = cost(y_hat, output_mini_batches[i]);
auto grads = compute_gradients(y_hat, output_mini_batches[i]);
auto cumulative_hidden_layer_w_grad = std::get<0>(grads);
auto output_w_grad = std::get<1>(grads);
if (!_network.empty() && v_hidden.empty()) { // Initing our tensor
v_hidden = alg.resize(v_hidden, cumulative_hidden_layer_w_grad);
}
if (v_output.empty()) {
v_output.resize(output_w_grad.size());
}
v_hidden = alg.addition(v_hidden, alg.exponentiate(cumulative_hidden_layer_w_grad, 2));
v_output = alg.addition(v_output, alg.exponentiate(output_w_grad, 2));
std::vector<std::vector<std::vector<real_t>>> hidden_layer_updations = alg.scalarMultiply(learning_rate / _n, alg.elementWiseDivision(cumulative_hidden_layer_w_grad, alg.scalarAdd(e, alg.sqrt(v_hidden))));
std::vector<real_t> output_layer_updation = alg.scalarMultiply(learning_rate / _n, alg.elementWiseDivision(output_w_grad, alg.scalarAdd(e, alg.sqrt(v_output))));
update_parameters(hidden_layer_updations, output_layer_updation, learning_rate); // subject to change. may want bias to have this matrix too.
y_hat = model_set_test(input_mini_batches[i]);
if (ui) {
print_ui(epoch, cost_prev, y_hat, output_mini_batches[i]);
}
}
epoch++;
if (epoch > max_epoch) {
break;
}
}
forward_pass();
}
void MLPPANN::adadelta(real_t learning_rate, int max_epoch, int mini_batch_size, real_t b1, real_t e, bool ui) {
MLPPCost mlpp_cost;
MLPPLinAlg alg;
real_t cost_prev = 0;
int epoch = 1;
real_t initial_learning_rate = learning_rate;
// Creating the mini-batches
int n_mini_batch = _n / mini_batch_size;
// always evaluate the result
// always do forward pass only ONCE at end.
auto batches = MLPPUtilities::createMiniBatches(_input_set, _output_set, n_mini_batch);
auto input_mini_batches = std::get<0>(batches);
auto output_mini_batches = std::get<1>(batches);
// Initializing necessary components for Adam.
std::vector<std::vector<std::vector<real_t>>> v_hidden;
std::vector<real_t> v_output;
while (true) {
learning_rate = apply_learning_rate_scheduler(initial_learning_rate, _decay_constant, epoch, _drop_rate);
for (int i = 0; i < n_mini_batch; i++) {
std::vector<real_t> y_hat = model_set_test(input_mini_batches[i]);
cost_prev = cost(y_hat, output_mini_batches[i]);
auto grads = compute_gradients(y_hat, output_mini_batches[i]);
auto cumulative_hidden_layer_w_grad = std::get<0>(grads);
auto output_w_grad = std::get<1>(grads);
if (!_network.empty() && v_hidden.empty()) { // Initing our tensor
v_hidden = alg.resize(v_hidden, cumulative_hidden_layer_w_grad);
}
if (v_output.empty()) {
v_output.resize(output_w_grad.size());
}
v_hidden = alg.addition(alg.scalarMultiply(1 - b1, v_hidden), alg.scalarMultiply(b1, alg.exponentiate(cumulative_hidden_layer_w_grad, 2)));
v_output = alg.addition(v_output, alg.exponentiate(output_w_grad, 2));
std::vector<std::vector<std::vector<real_t>>> hidden_layer_updations = alg.scalarMultiply(learning_rate / _n, alg.elementWiseDivision(cumulative_hidden_layer_w_grad, alg.scalarAdd(e, alg.sqrt(v_hidden))));
std::vector<real_t> output_layer_updation = alg.scalarMultiply(learning_rate / _n, alg.elementWiseDivision(output_w_grad, alg.scalarAdd(e, alg.sqrt(v_output))));
update_parameters(hidden_layer_updations, output_layer_updation, learning_rate); // subject to change. may want bias to have this matrix too.
y_hat = model_set_test(input_mini_batches[i]);
if (ui) {
print_ui(epoch, cost_prev, y_hat, output_mini_batches[i]);
}
}
epoch++;
if (epoch > max_epoch) {
break;
}
}
forward_pass();
}
void MLPPANN::adam(real_t learning_rate, int max_epoch, int mini_batch_size, real_t b1, real_t b2, real_t e, bool ui) {
MLPPCost mlpp_cost;
MLPPLinAlg alg;
real_t cost_prev = 0;
int epoch = 1;
real_t initial_learning_rate = learning_rate;
// Creating the mini-batches
int n_mini_batch = _n / mini_batch_size;
// always evaluate the result
// always do forward pass only ONCE at end.
auto batches = MLPPUtilities::createMiniBatches(_input_set, _output_set, n_mini_batch);
auto input_mini_batches = std::get<0>(batches);
auto output_mini_batches = std::get<1>(batches);
// Initializing necessary components for Adam.
std::vector<std::vector<std::vector<real_t>>> m_hidden;
std::vector<std::vector<std::vector<real_t>>> v_hidden;
std::vector<real_t> m_output;
std::vector<real_t> v_output;
while (true) {
learning_rate = apply_learning_rate_scheduler(initial_learning_rate, _decay_constant, epoch, _drop_rate);
for (int i = 0; i < n_mini_batch; i++) {
std::vector<real_t> y_hat = model_set_test(input_mini_batches[i]);
cost_prev = cost(y_hat, output_mini_batches[i]);
auto grads = compute_gradients(y_hat, output_mini_batches[i]);
auto cumulative_hidden_layer_w_grad = std::get<0>(grads);
auto output_w_grad = std::get<1>(grads);
if (!_network.empty() && m_hidden.empty() && v_hidden.empty()) { // Initing our tensor
m_hidden = alg.resize(m_hidden, cumulative_hidden_layer_w_grad);
v_hidden = alg.resize(v_hidden, cumulative_hidden_layer_w_grad);
}
if (m_output.empty() && v_output.empty()) {
m_output.resize(output_w_grad.size());
v_output.resize(output_w_grad.size());
}
m_hidden = alg.addition(alg.scalarMultiply(b1, m_hidden), alg.scalarMultiply(1 - b1, cumulative_hidden_layer_w_grad));
v_hidden = alg.addition(alg.scalarMultiply(b2, v_hidden), alg.scalarMultiply(1 - b2, alg.exponentiate(cumulative_hidden_layer_w_grad, 2)));
m_output = alg.addition(alg.scalarMultiply(b1, m_output), alg.scalarMultiply(1 - b1, output_w_grad));
v_output = alg.addition(alg.scalarMultiply(b2, v_output), alg.scalarMultiply(1 - b2, alg.exponentiate(output_w_grad, 2)));
std::vector<std::vector<std::vector<real_t>>> m_hidden_hat = alg.scalarMultiply(1 / (1 - std::pow(b1, epoch)), m_hidden);
std::vector<std::vector<std::vector<real_t>>> v_hidden_hat = alg.scalarMultiply(1 / (1 - std::pow(b2, epoch)), v_hidden);
std::vector<real_t> m_output_hat = alg.scalarMultiply(1 / (1 - std::pow(b1, epoch)), m_output);
std::vector<real_t> v_output_hat = alg.scalarMultiply(1 / (1 - std::pow(b2, epoch)), v_output);
std::vector<std::vector<std::vector<real_t>>> hidden_layer_updations = alg.scalarMultiply(learning_rate / _n, alg.elementWiseDivision(m_hidden_hat, alg.scalarAdd(e, alg.sqrt(v_hidden_hat))));
std::vector<real_t> output_layer_updation = alg.scalarMultiply(learning_rate / _n, alg.elementWiseDivision(m_output_hat, alg.scalarAdd(e, alg.sqrt(v_output_hat))));
update_parameters(hidden_layer_updations, output_layer_updation, learning_rate); // subject to change. may want bias to have this matrix too.
y_hat = model_set_test(input_mini_batches[i]);
if (ui) {
print_ui(epoch, cost_prev, y_hat, output_mini_batches[i]);
}
}
epoch++;
if (epoch > max_epoch) {
break;
}
}
forward_pass();
}
void MLPPANN::adamax(real_t learning_rate, int max_epoch, int mini_batch_size, real_t b1, real_t b2, real_t e, bool ui) {
MLPPCost mlpp_cost;
MLPPLinAlg alg;
real_t cost_prev = 0;
int epoch = 1;
real_t initial_learning_rate = learning_rate;
// Creating the mini-batches
int n_mini_batch = _n / mini_batch_size;
// always evaluate the result
// always do forward pass only ONCE at end.
auto batches = MLPPUtilities::createMiniBatches(_input_set, _output_set, n_mini_batch);
auto input_mini_batches = std::get<0>(batches);
auto output_mini_batches = std::get<1>(batches);
// Initializing necessary components for Adam.
std::vector<std::vector<std::vector<real_t>>> m_hidden;
std::vector<std::vector<std::vector<real_t>>> u_hidden;
std::vector<real_t> m_output;
std::vector<real_t> u_output;
while (true) {
learning_rate = apply_learning_rate_scheduler(initial_learning_rate, _decay_constant, epoch, _drop_rate);
for (int i = 0; i < n_mini_batch; i++) {
std::vector<real_t> y_hat = model_set_test(input_mini_batches[i]);
cost_prev = cost(y_hat, output_mini_batches[i]);
auto grads = compute_gradients(y_hat, output_mini_batches[i]);
auto cumulative_hidden_layer_w_grad = std::get<0>(grads);
auto output_w_grad = std::get<1>(grads);
if (!_network.empty() && m_hidden.empty() && u_hidden.empty()) { // Initing our tensor
m_hidden = alg.resize(m_hidden, cumulative_hidden_layer_w_grad);
u_hidden = alg.resize(u_hidden, cumulative_hidden_layer_w_grad);
}
if (m_output.empty() && u_output.empty()) {
m_output.resize(output_w_grad.size());
u_output.resize(output_w_grad.size());
}
m_hidden = alg.addition(alg.scalarMultiply(b1, m_hidden), alg.scalarMultiply(1 - b1, cumulative_hidden_layer_w_grad));
u_hidden = alg.max(alg.scalarMultiply(b2, u_hidden), alg.abs(cumulative_hidden_layer_w_grad));
m_output = alg.addition(alg.scalarMultiply(b1, m_output), alg.scalarMultiply(1 - b1, output_w_grad));
u_output = alg.max(alg.scalarMultiply(b2, u_output), alg.abs(output_w_grad));
std::vector<std::vector<std::vector<real_t>>> m_hidden_hat = alg.scalarMultiply(1 / (1 - std::pow(b1, epoch)), m_hidden);
std::vector<real_t> m_output_hat = alg.scalarMultiply(1 / (1 - std::pow(b1, epoch)), m_output);
std::vector<std::vector<std::vector<real_t>>> hidden_layer_updations = alg.scalarMultiply(learning_rate / _n, alg.elementWiseDivision(m_hidden_hat, alg.scalarAdd(e, u_hidden)));
std::vector<real_t> output_layer_updation = alg.scalarMultiply(learning_rate / _n, alg.elementWiseDivision(m_output_hat, alg.scalarAdd(e, u_output)));
update_parameters(hidden_layer_updations, output_layer_updation, learning_rate); // subject to change. may want bias to have this matrix too.
y_hat = model_set_test(input_mini_batches[i]);
if (ui) {
print_ui(epoch, cost_prev, y_hat, output_mini_batches[i]);
}
}
epoch++;
if (epoch > max_epoch) {
break;
}
}
forward_pass();
}
void MLPPANN::nadam(real_t learning_rate, int max_epoch, int mini_batch_size, real_t b1, real_t b2, real_t e, bool ui) {
MLPPCost mlpp_cost;
MLPPLinAlg alg;
real_t cost_prev = 0;
int epoch = 1;
real_t initial_learning_rate = learning_rate;
// Creating the mini-batches
int n_mini_batch = _n / mini_batch_size;
// always evaluate the result
// always do forward pass only ONCE at end.
auto batches = MLPPUtilities::createMiniBatches(_input_set, _output_set, n_mini_batch);
auto input_mini_batches = std::get<0>(batches);
auto output_mini_batches = std::get<1>(batches);
// Initializing necessary components for Adam.
std::vector<std::vector<std::vector<real_t>>> m_hidden;
std::vector<std::vector<std::vector<real_t>>> v_hidden;
std::vector<real_t> m_output;
std::vector<real_t> v_output;
while (true) {
learning_rate = apply_learning_rate_scheduler(initial_learning_rate, _decay_constant, epoch, _drop_rate);
for (int i = 0; i < n_mini_batch; i++) {
std::vector<real_t> y_hat = model_set_test(input_mini_batches[i]);
cost_prev = cost(y_hat, output_mini_batches[i]);
auto grads = compute_gradients(y_hat, output_mini_batches[i]);
auto cumulative_hidden_layer_w_grad = std::get<0>(grads);
auto output_w_grad = std::get<1>(grads);
if (!_network.empty() && m_hidden.empty() && v_hidden.empty()) { // Initing our tensor
m_hidden = alg.resize(m_hidden, cumulative_hidden_layer_w_grad);
v_hidden = alg.resize(v_hidden, cumulative_hidden_layer_w_grad);
}
if (m_output.empty() && v_output.empty()) {
m_output.resize(output_w_grad.size());
v_output.resize(output_w_grad.size());
}
m_hidden = alg.addition(alg.scalarMultiply(b1, m_hidden), alg.scalarMultiply(1 - b1, cumulative_hidden_layer_w_grad));
v_hidden = alg.addition(alg.scalarMultiply(b2, v_hidden), alg.scalarMultiply(1 - b2, alg.exponentiate(cumulative_hidden_layer_w_grad, 2)));
m_output = alg.addition(alg.scalarMultiply(b1, m_output), alg.scalarMultiply(1 - b1, output_w_grad));
v_output = alg.addition(alg.scalarMultiply(b2, v_output), alg.scalarMultiply(1 - b2, alg.exponentiate(output_w_grad, 2)));
std::vector<std::vector<std::vector<real_t>>> m_hidden_hat = alg.scalarMultiply(1 / (1 - std::pow(b1, epoch)), m_hidden);
std::vector<std::vector<std::vector<real_t>>> v_hidden_hat = alg.scalarMultiply(1 / (1 - std::pow(b2, epoch)), v_hidden);
std::vector<std::vector<std::vector<real_t>>> m_hidden_final = alg.addition(alg.scalarMultiply(b1, m_hidden_hat), alg.scalarMultiply((1 - b1) / (1 - std::pow(b1, epoch)), cumulative_hidden_layer_w_grad));
std::vector<real_t> m_output_hat = alg.scalarMultiply(1 / (1 - std::pow(b1, epoch)), m_output);
std::vector<real_t> v_output_hat = alg.scalarMultiply(1 / (1 - std::pow(b2, epoch)), v_output);
std::vector<real_t> m_output_final = alg.addition(alg.scalarMultiply(b1, m_output_hat), alg.scalarMultiply((1 - b1) / (1 - std::pow(b1, epoch)), output_w_grad));
std::vector<std::vector<std::vector<real_t>>> hidden_layer_updations = alg.scalarMultiply(learning_rate / _n, alg.elementWiseDivision(m_hidden_final, alg.scalarAdd(e, alg.sqrt(v_hidden_hat))));
std::vector<real_t> output_layer_updation = alg.scalarMultiply(learning_rate / _n, alg.elementWiseDivision(m_output_final, alg.scalarAdd(e, alg.sqrt(v_output_hat))));
update_parameters(hidden_layer_updations, output_layer_updation, learning_rate); // subject to change. may want bias to have this matrix too.
y_hat = model_set_test(input_mini_batches[i]);
if (ui) {
print_ui(epoch, cost_prev, y_hat, output_mini_batches[i]);
}
}
epoch++;
if (epoch > max_epoch) {
break;
}
}
forward_pass();
}
void MLPPANN::amsgrad(real_t learning_rate, int max_epoch, int mini_batch_size, real_t b1, real_t b2, real_t e, bool ui) {
MLPPCost mlpp_cost;
MLPPLinAlg alg;
real_t cost_prev = 0;
int epoch = 1;
real_t initial_learning_rate = learning_rate;
// Creating the mini-batches
int n_mini_batch = _n / mini_batch_size;
// always evaluate the result
// always do forward pass only ONCE at end.
auto batches = MLPPUtilities::createMiniBatches(_input_set, _output_set, n_mini_batch);
auto input_mini_batches = std::get<0>(batches);
auto output_mini_batches = std::get<1>(batches);
// Initializing necessary components for Adam.
std::vector<std::vector<std::vector<real_t>>> m_hidden;
std::vector<std::vector<std::vector<real_t>>> v_hidden;
std::vector<std::vector<std::vector<real_t>>> v_hidden_hat;
std::vector<real_t> m_output;
std::vector<real_t> v_output;
std::vector<real_t> v_output_hat;
while (true) {
learning_rate = apply_learning_rate_scheduler(initial_learning_rate, _decay_constant, epoch, _drop_rate);
for (int i = 0; i < n_mini_batch; i++) {
std::vector<real_t> y_hat = model_set_test(input_mini_batches[i]);
cost_prev = cost(y_hat, output_mini_batches[i]);
auto grads = compute_gradients(y_hat, output_mini_batches[i]);
auto cumulative_hidden_layer_w_grad = std::get<0>(grads);
auto output_w_grad = std::get<1>(grads);
if (!_network.empty() && m_hidden.empty() && v_hidden.empty()) { // Initing our tensor
m_hidden = alg.resize(m_hidden, cumulative_hidden_layer_w_grad);
v_hidden = alg.resize(v_hidden, cumulative_hidden_layer_w_grad);
v_hidden_hat = alg.resize(v_hidden_hat, cumulative_hidden_layer_w_grad);
}
if (m_output.empty() && v_output.empty()) {
m_output.resize(output_w_grad.size());
v_output.resize(output_w_grad.size());
v_output_hat.resize(output_w_grad.size());
}
m_hidden = alg.addition(alg.scalarMultiply(b1, m_hidden), alg.scalarMultiply(1 - b1, cumulative_hidden_layer_w_grad));
v_hidden = alg.addition(alg.scalarMultiply(b2, v_hidden), alg.scalarMultiply(1 - b2, alg.exponentiate(cumulative_hidden_layer_w_grad, 2)));
m_output = alg.addition(alg.scalarMultiply(b1, m_output), alg.scalarMultiply(1 - b1, output_w_grad));
v_output = alg.addition(alg.scalarMultiply(b2, v_output), alg.scalarMultiply(1 - b2, alg.exponentiate(output_w_grad, 2)));
v_hidden_hat = alg.max(v_hidden_hat, v_hidden);
v_output_hat = alg.max(v_output_hat, v_output);
std::vector<std::vector<std::vector<real_t>>> hidden_layer_updations = alg.scalarMultiply(learning_rate / _n, alg.elementWiseDivision(m_hidden, alg.scalarAdd(e, alg.sqrt(v_hidden_hat))));
std::vector<real_t> output_layer_updation = alg.scalarMultiply(learning_rate / _n, alg.elementWiseDivision(m_output, alg.scalarAdd(e, alg.sqrt(v_output_hat))));
update_parameters(hidden_layer_updations, output_layer_updation, learning_rate); // subject to change. may want bias to have this matrix too.
y_hat = model_set_test(input_mini_batches[i]);
if (ui) {
print_ui(epoch, cost_prev, y_hat, output_mini_batches[i]);
}
}
epoch++;
if (epoch > max_epoch) {
break;
}
}
forward_pass();
}
real_t MLPPANN::score() {
MLPPUtilities util;
forward_pass();
return util.performance(_y_hat, _output_set);
}
void MLPPANN::save(std::string fileName) {
MLPPUtilities util;
if (!_network.empty()) {
util.saveParameters(fileName, _network[0].weights, _network[0].bias, false, 1);
for (uint32_t i = 1; i < _network.size(); i++) {
util.saveParameters(fileName, _network[i].weights, _network[i].bias, true, i + 1);
}
util.saveParameters(fileName, _output_layer->weights, _output_layer->bias, true, _network.size() + 1);
} else {
util.saveParameters(fileName, _output_layer->weights, _output_layer->bias, false, _network.size() + 1);
}
}
void MLPPANN::set_learning_rate_scheduler(std::string type, real_t decay_constant) {
_lr_scheduler = type;
_decay_constant = decay_constant;
}
void MLPPANN::set_learning_rate_scheduler_drop(std::string type, real_t decay_constant, real_t drop_rate) {
_lr_scheduler = type;
_decay_constant = decay_constant;
_drop_rate = drop_rate;
}
// https://en.wikipedia.org/wiki/Learning_rate
// Learning Rate Decay (C2W2L09) - Andrew Ng - Deep Learning Specialization
real_t MLPPANN::apply_learning_rate_scheduler(real_t learning_rate, real_t decay_constant, real_t epoch, real_t drop_rate) {
if (_lr_scheduler == "Time") {
return learning_rate / (1 + decay_constant * epoch);
} else if (_lr_scheduler == "Epoch") {
return learning_rate * (decay_constant / std::sqrt(epoch));
} else if (_lr_scheduler == "Step") {
return learning_rate * std::pow(decay_constant, int((1 + epoch) / drop_rate)); // Utilizing an explicit int conversion implicitly takes the floor.
} else if (_lr_scheduler == "Exponential") {
return learning_rate * std::exp(-decay_constant * epoch);
}
return learning_rate;
}
void MLPPANN::add_layer(int n_hidden, std::string activation, std::string weightInit, std::string reg, real_t lambda, real_t alpha) {
if (_network.empty()) {
_network.push_back(MLPPOldHiddenLayer(n_hidden, activation, _input_set, weightInit, reg, lambda, alpha));
_network[0].forwardPass();
} else {
_network.push_back(MLPPOldHiddenLayer(n_hidden, activation, _network[_network.size() - 1].a, weightInit, reg, lambda, alpha));
_network[_network.size() - 1].forwardPass();
}
}
void MLPPANN::add_output_layer(std::string activation, std::string loss, std::string weightInit, std::string reg, real_t lambda, real_t alpha) {
if (!_network.empty()) {
_output_layer = new MLPPOldOutputLayer(_network[_network.size() - 1].n_hidden, activation, loss, _network[_network.size() - 1].a, weightInit, reg, lambda, alpha);
} else {
_output_layer = new MLPPOldOutputLayer(_k, activation, loss, _input_set, weightInit, reg, lambda, alpha);
}
}
MLPPANN::MLPPANN(std::vector<std::vector<real_t>> p_input_set, std::vector<real_t> p_output_set) {
_input_set = p_input_set;
_output_set = p_output_set;
_n = _input_set.size();
_k = _input_set[0].size();
_lr_scheduler = "None";
_decay_constant = 0;
_drop_rate = 0;
}
MLPPANN::MLPPANN() {
}
MLPPANN::~MLPPANN() {
delete _output_layer;
}
real_t MLPPANN::cost(std::vector<real_t> y_hat, std::vector<real_t> y) {
MLPPReg regularization;
MLPPCost mlpp_cost;
real_t totalRegTerm = 0;
auto cost_function = _output_layer->cost_map[_output_layer->cost];
if (!_network.empty()) {
for (uint32_t i = 0; i < _network.size() - 1; i++) {
totalRegTerm += regularization.regTerm(_network[i].weights, _network[i].lambda, _network[i].alpha, _network[i].reg);
}
}
return (mlpp_cost.*cost_function)(y_hat, y) + totalRegTerm + regularization.regTerm(_output_layer->weights, _output_layer->lambda, _output_layer->alpha, _output_layer->reg);
}
void MLPPANN::forward_pass() {
if (!_network.empty()) {
_network[0].input = _input_set;
_network[0].forwardPass();
for (uint32_t i = 1; i < _network.size(); i++) {
_network[i].input = _network[i - 1].a;
_network[i].forwardPass();
}
_output_layer->input = _network[_network.size() - 1].a;
} else {
_output_layer->input = _input_set;
}
_output_layer->forwardPass();
_y_hat = _output_layer->a;
}
void MLPPANN::update_parameters(std::vector<std::vector<std::vector<real_t>>> hidden_layer_updations, std::vector<real_t> output_layer_updation, real_t learning_rate) {
MLPPLinAlg alg;
_output_layer->weights = alg.subtraction(_output_layer->weights, output_layer_updation);
_output_layer->bias -= learning_rate * alg.sum_elements(_output_layer->delta) / _n;
if (!_network.empty()) {
_network[_network.size() - 1].weights = alg.subtraction(_network[_network.size() - 1].weights, hidden_layer_updations[0]);
_network[_network.size() - 1].bias = alg.subtractMatrixRows(_network[_network.size() - 1].bias, alg.scalarMultiply(learning_rate / _n, _network[_network.size() - 1].delta));
for (int i = _network.size() - 2; i >= 0; i--) {
_network[i].weights = alg.subtraction(_network[i].weights, hidden_layer_updations[(_network.size() - 2) - i + 1]);
_network[i].bias = alg.subtractMatrixRows(_network[i].bias, alg.scalarMultiply(learning_rate / _n, _network[i].delta));
}
}
}
std::tuple<std::vector<std::vector<std::vector<real_t>>>, std::vector<real_t>> MLPPANN::compute_gradients(std::vector<real_t> y_hat, std::vector<real_t> _output_set) {
// std::cout << "BEGIN" << std::endl;
MLPPCost mlpp_cost;
MLPPActivation avn;
MLPPLinAlg alg;
MLPPReg regularization;
std::vector<std::vector<std::vector<real_t>>> cumulative_hidden_layer_w_grad; // Tensor containing ALL hidden grads.
auto costDeriv = _output_layer->costDeriv_map[_output_layer->cost];
auto outputAvn = _output_layer->activation_map[_output_layer->activation];
_output_layer->delta = alg.hadamard_product((mlpp_cost.*costDeriv)(y_hat, _output_set), (avn.*outputAvn)(_output_layer->z, 1));
std::vector<real_t> output_w_grad = alg.mat_vec_mult(alg.transpose(_output_layer->input), _output_layer->delta);
output_w_grad = alg.addition(output_w_grad, regularization.regDerivTerm(_output_layer->weights, _output_layer->lambda, _output_layer->alpha, _output_layer->reg));
if (!_network.empty()) {
auto hiddenLayerAvn = _network[_network.size() - 1].activation_map[_network[_network.size() - 1].activation];
_network[_network.size() - 1].delta = alg.hadamard_product(alg.outerProduct(_output_layer->delta, _output_layer->weights), (avn.*hiddenLayerAvn)(_network[_network.size() - 1].z, 1));
std::vector<std::vector<real_t>> hiddenLayerWGrad = alg.matmult(alg.transpose(_network[_network.size() - 1].input), _network[_network.size() - 1].delta);
cumulative_hidden_layer_w_grad.push_back(alg.addition(hiddenLayerWGrad, regularization.regDerivTerm(_network[_network.size() - 1].weights, _network[_network.size() - 1].lambda, _network[_network.size() - 1].alpha, _network[_network.size() - 1].reg))); // Adding to our cumulative hidden layer grads. Maintain reg terms as well.
for (int i = _network.size() - 2; i >= 0; i--) {
hiddenLayerAvn = _network[i].activation_map[_network[i].activation];
_network[i].delta = alg.hadamard_product(alg.matmult(_network[i + 1].delta, alg.transpose(_network[i + 1].weights)), (avn.*hiddenLayerAvn)(_network[i].z, 1));
hiddenLayerWGrad = alg.matmult(alg.transpose(_network[i].input), _network[i].delta);
cumulative_hidden_layer_w_grad.push_back(alg.addition(hiddenLayerWGrad, regularization.regDerivTerm(_network[i].weights, _network[i].lambda, _network[i].alpha, _network[i].reg))); // Adding to our cumulative hidden layer grads. Maintain reg terms as well.
}
}
return { cumulative_hidden_layer_w_grad, output_w_grad };
}
void MLPPANN::print_ui(int epoch, real_t cost_prev, std::vector<real_t> y_hat, std::vector<real_t> p_output_set) {
MLPPUtilities::CostInfo(epoch, cost_prev, cost(y_hat, p_output_set));
std::cout << "Layer " << _network.size() + 1 << ": " << std::endl;
MLPPUtilities::UI(_output_layer->weights, _output_layer->bias);
if (!_network.empty()) {
for (int i = _network.size() - 1; i >= 0; i--) {
std::cout << "Layer " << i + 1 << ": " << std::endl;
MLPPUtilities::UI(_network[i].weights, _network[i].bias);
}
}
}
void MLPPANN::_bind_methods() {
}