mirror of https://github.com/Relintai/pmlpp.git
819 lines
32 KiB
C++
819 lines
32 KiB
C++
//
|
|
// ANN.cpp
|
|
//
|
|
// Created by Marc Melikyan on 11/4/20.
|
|
//
|
|
|
|
#include "ann_old.h"
|
|
#include "../activation/activation_old.h"
|
|
#include "../cost/cost.h"
|
|
#include "../lin_alg/lin_alg_old.h"
|
|
#include "../regularization/reg.h"
|
|
#include "../utilities/utilities.h"
|
|
|
|
#include <cmath>
|
|
#include <iostream>
|
|
#include <random>
|
|
|
|
MLPPANNOld::MLPPANNOld(std::vector<std::vector<real_t>> p_inputSet, std::vector<real_t> p_outputSet) {
|
|
inputSet = p_inputSet;
|
|
outputSet = p_outputSet;
|
|
|
|
n = inputSet.size();
|
|
k = inputSet[0].size();
|
|
lrScheduler = "None";
|
|
decayConstant = 0;
|
|
dropRate = 0;
|
|
}
|
|
|
|
MLPPANNOld::~MLPPANNOld() {
|
|
delete outputLayer;
|
|
}
|
|
|
|
std::vector<real_t> MLPPANNOld::modelSetTest(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();
|
|
}
|
|
outputLayer->input = network[network.size() - 1].a;
|
|
} else {
|
|
outputLayer->input = X;
|
|
}
|
|
outputLayer->forwardPass();
|
|
return outputLayer->a;
|
|
}
|
|
|
|
real_t MLPPANNOld::modelTest(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);
|
|
}
|
|
outputLayer->Test(network[network.size() - 1].a_test);
|
|
} else {
|
|
outputLayer->Test(x);
|
|
}
|
|
return outputLayer->a_test;
|
|
}
|
|
|
|
void MLPPANNOld::gradientDescent(real_t learning_rate, int max_epoch, bool UI) {
|
|
class MLPPCost cost;
|
|
MLPPLinAlgOld alg;
|
|
real_t cost_prev = 0;
|
|
int epoch = 1;
|
|
forwardPass();
|
|
real_t initial_learning_rate = learning_rate;
|
|
|
|
alg.printMatrix(network[network.size() - 1].weights);
|
|
while (true) {
|
|
learning_rate = applyLearningRateScheduler(initial_learning_rate, decayConstant, epoch, dropRate);
|
|
cost_prev = Cost(y_hat, outputSet);
|
|
|
|
auto grads = computeGradients(y_hat, outputSet);
|
|
auto cumulativeHiddenLayerWGrad = std::get<0>(grads);
|
|
auto outputWGrad = std::get<1>(grads);
|
|
|
|
cumulativeHiddenLayerWGrad = alg.scalarMultiply(learning_rate / n, cumulativeHiddenLayerWGrad);
|
|
outputWGrad = alg.scalarMultiply(learning_rate / n, outputWGrad);
|
|
updateParameters(cumulativeHiddenLayerWGrad, outputWGrad, learning_rate); // subject to change. may want bias to have this matrix too.
|
|
|
|
std::cout << learning_rate << std::endl;
|
|
|
|
forwardPass();
|
|
|
|
if (UI) {
|
|
MLPPANNOld::UI(epoch, cost_prev, y_hat, outputSet);
|
|
}
|
|
|
|
epoch++;
|
|
if (epoch > max_epoch) {
|
|
break;
|
|
}
|
|
}
|
|
}
|
|
|
|
void MLPPANNOld::SGD(real_t learning_rate, int max_epoch, bool UI) {
|
|
class MLPPCost cost;
|
|
MLPPLinAlgOld alg;
|
|
|
|
real_t cost_prev = 0;
|
|
int epoch = 1;
|
|
real_t initial_learning_rate = learning_rate;
|
|
|
|
while (true) {
|
|
learning_rate = applyLearningRateScheduler(initial_learning_rate, decayConstant, epoch, dropRate);
|
|
|
|
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 = modelSetTest({ inputSet[outputIndex] });
|
|
cost_prev = Cost({ y_hat }, { outputSet[outputIndex] });
|
|
|
|
auto grads = computeGradients(y_hat, { outputSet[outputIndex] });
|
|
auto cumulativeHiddenLayerWGrad = std::get<0>(grads);
|
|
auto outputWGrad = std::get<1>(grads);
|
|
|
|
cumulativeHiddenLayerWGrad = alg.scalarMultiply(learning_rate / n, cumulativeHiddenLayerWGrad);
|
|
outputWGrad = alg.scalarMultiply(learning_rate / n, outputWGrad);
|
|
|
|
updateParameters(cumulativeHiddenLayerWGrad, outputWGrad, learning_rate); // subject to change. may want bias to have this matrix too.
|
|
y_hat = modelSetTest({ inputSet[outputIndex] });
|
|
|
|
if (UI) {
|
|
MLPPANNOld::UI(epoch, cost_prev, y_hat, { outputSet[outputIndex] });
|
|
}
|
|
|
|
epoch++;
|
|
if (epoch > max_epoch) {
|
|
break;
|
|
}
|
|
}
|
|
forwardPass();
|
|
}
|
|
|
|
void MLPPANNOld::MBGD(real_t learning_rate, int max_epoch, int mini_batch_size, bool UI) {
|
|
class MLPPCost cost;
|
|
MLPPLinAlgOld 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(inputSet, outputSet, n_mini_batch);
|
|
auto inputMiniBatches = std::get<0>(batches);
|
|
auto outputMiniBatches = std::get<1>(batches);
|
|
|
|
while (true) {
|
|
learning_rate = applyLearningRateScheduler(initial_learning_rate, decayConstant, epoch, dropRate);
|
|
for (int i = 0; i < n_mini_batch; i++) {
|
|
std::vector<real_t> y_hat = modelSetTest(inputMiniBatches[i]);
|
|
cost_prev = Cost(y_hat, outputMiniBatches[i]);
|
|
|
|
auto grads = computeGradients(y_hat, outputMiniBatches[i]);
|
|
auto cumulativeHiddenLayerWGrad = std::get<0>(grads);
|
|
auto outputWGrad = std::get<1>(grads);
|
|
|
|
cumulativeHiddenLayerWGrad = alg.scalarMultiply(learning_rate / n, cumulativeHiddenLayerWGrad);
|
|
outputWGrad = alg.scalarMultiply(learning_rate / n, outputWGrad);
|
|
|
|
updateParameters(cumulativeHiddenLayerWGrad, outputWGrad, learning_rate); // subject to change. may want bias to have this matrix too.
|
|
y_hat = modelSetTest(inputMiniBatches[i]);
|
|
|
|
if (UI) {
|
|
MLPPANNOld::UI(epoch, cost_prev, y_hat, outputMiniBatches[i]);
|
|
}
|
|
}
|
|
epoch++;
|
|
if (epoch > max_epoch) {
|
|
break;
|
|
}
|
|
}
|
|
forwardPass();
|
|
}
|
|
|
|
void MLPPANNOld::Momentum(real_t learning_rate, int max_epoch, int mini_batch_size, real_t gamma, bool NAG, bool UI) {
|
|
class MLPPCost cost;
|
|
MLPPLinAlgOld 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(inputSet, outputSet, n_mini_batch);
|
|
auto inputMiniBatches = std::get<0>(batches);
|
|
auto outputMiniBatches = 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 = applyLearningRateScheduler(initial_learning_rate, decayConstant, epoch, dropRate);
|
|
for (int i = 0; i < n_mini_batch; i++) {
|
|
std::vector<real_t> y_hat = modelSetTest(inputMiniBatches[i]);
|
|
cost_prev = Cost(y_hat, outputMiniBatches[i]);
|
|
|
|
auto grads = computeGradients(y_hat, outputMiniBatches[i]);
|
|
auto cumulativeHiddenLayerWGrad = std::get<0>(grads);
|
|
auto outputWGrad = std::get<1>(grads);
|
|
|
|
if (!network.empty() && v_hidden.empty()) { // Initing our tensor
|
|
v_hidden = alg.resize(v_hidden, cumulativeHiddenLayerWGrad);
|
|
}
|
|
|
|
if (v_output.empty()) {
|
|
v_output.resize(outputWGrad.size());
|
|
}
|
|
|
|
if (NAG) { // "Aposterori" calculation
|
|
updateParameters(v_hidden, v_output, 0); // DON'T update bias.
|
|
}
|
|
|
|
v_hidden = alg.addition(alg.scalarMultiply(gamma, v_hidden), alg.scalarMultiply(learning_rate / n, cumulativeHiddenLayerWGrad));
|
|
|
|
v_output = alg.addition(alg.scalarMultiply(gamma, v_output), alg.scalarMultiply(learning_rate / n, outputWGrad));
|
|
|
|
updateParameters(v_hidden, v_output, learning_rate); // subject to change. may want bias to have this matrix too.
|
|
y_hat = modelSetTest(inputMiniBatches[i]);
|
|
|
|
if (UI) {
|
|
MLPPANNOld::UI(epoch, cost_prev, y_hat, outputMiniBatches[i]);
|
|
}
|
|
}
|
|
epoch++;
|
|
if (epoch > max_epoch) {
|
|
break;
|
|
}
|
|
}
|
|
forwardPass();
|
|
}
|
|
|
|
void MLPPANNOld::Adagrad(real_t learning_rate, int max_epoch, int mini_batch_size, real_t e, bool UI) {
|
|
class MLPPCost cost;
|
|
MLPPLinAlgOld 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(inputSet, outputSet, n_mini_batch);
|
|
auto inputMiniBatches = std::get<0>(batches);
|
|
auto outputMiniBatches = 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 = applyLearningRateScheduler(initial_learning_rate, decayConstant, epoch, dropRate);
|
|
for (int i = 0; i < n_mini_batch; i++) {
|
|
std::vector<real_t> y_hat = modelSetTest(inputMiniBatches[i]);
|
|
cost_prev = Cost(y_hat, outputMiniBatches[i]);
|
|
|
|
auto grads = computeGradients(y_hat, outputMiniBatches[i]);
|
|
auto cumulativeHiddenLayerWGrad = std::get<0>(grads);
|
|
auto outputWGrad = std::get<1>(grads);
|
|
|
|
if (!network.empty() && v_hidden.empty()) { // Initing our tensor
|
|
v_hidden = alg.resize(v_hidden, cumulativeHiddenLayerWGrad);
|
|
}
|
|
|
|
if (v_output.empty()) {
|
|
v_output.resize(outputWGrad.size());
|
|
}
|
|
|
|
v_hidden = alg.addition(v_hidden, alg.exponentiate(cumulativeHiddenLayerWGrad, 2));
|
|
|
|
v_output = alg.addition(v_output, alg.exponentiate(outputWGrad, 2));
|
|
|
|
std::vector<std::vector<std::vector<real_t>>> hiddenLayerUpdations = alg.scalarMultiply(learning_rate / n, alg.elementWiseDivision(cumulativeHiddenLayerWGrad, alg.scalarAdd(e, alg.sqrt(v_hidden))));
|
|
std::vector<real_t> outputLayerUpdation = alg.scalarMultiply(learning_rate / n, alg.elementWiseDivision(outputWGrad, alg.scalarAdd(e, alg.sqrt(v_output))));
|
|
|
|
updateParameters(hiddenLayerUpdations, outputLayerUpdation, learning_rate); // subject to change. may want bias to have this matrix too.
|
|
y_hat = modelSetTest(inputMiniBatches[i]);
|
|
|
|
if (UI) {
|
|
MLPPANNOld::UI(epoch, cost_prev, y_hat, outputMiniBatches[i]);
|
|
}
|
|
}
|
|
epoch++;
|
|
if (epoch > max_epoch) {
|
|
break;
|
|
}
|
|
}
|
|
forwardPass();
|
|
}
|
|
|
|
void MLPPANNOld::Adadelta(real_t learning_rate, int max_epoch, int mini_batch_size, real_t b1, real_t e, bool UI) {
|
|
class MLPPCost cost;
|
|
MLPPLinAlgOld 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(inputSet, outputSet, n_mini_batch);
|
|
auto inputMiniBatches = std::get<0>(batches);
|
|
auto outputMiniBatches = 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 = applyLearningRateScheduler(initial_learning_rate, decayConstant, epoch, dropRate);
|
|
for (int i = 0; i < n_mini_batch; i++) {
|
|
std::vector<real_t> y_hat = modelSetTest(inputMiniBatches[i]);
|
|
cost_prev = Cost(y_hat, outputMiniBatches[i]);
|
|
|
|
auto grads = computeGradients(y_hat, outputMiniBatches[i]);
|
|
auto cumulativeHiddenLayerWGrad = std::get<0>(grads);
|
|
auto outputWGrad = std::get<1>(grads);
|
|
|
|
if (!network.empty() && v_hidden.empty()) { // Initing our tensor
|
|
v_hidden = alg.resize(v_hidden, cumulativeHiddenLayerWGrad);
|
|
}
|
|
|
|
if (v_output.empty()) {
|
|
v_output.resize(outputWGrad.size());
|
|
}
|
|
|
|
v_hidden = alg.addition(alg.scalarMultiply(1 - b1, v_hidden), alg.scalarMultiply(b1, alg.exponentiate(cumulativeHiddenLayerWGrad, 2)));
|
|
|
|
v_output = alg.addition(v_output, alg.exponentiate(outputWGrad, 2));
|
|
|
|
std::vector<std::vector<std::vector<real_t>>> hiddenLayerUpdations = alg.scalarMultiply(learning_rate / n, alg.elementWiseDivision(cumulativeHiddenLayerWGrad, alg.scalarAdd(e, alg.sqrt(v_hidden))));
|
|
std::vector<real_t> outputLayerUpdation = alg.scalarMultiply(learning_rate / n, alg.elementWiseDivision(outputWGrad, alg.scalarAdd(e, alg.sqrt(v_output))));
|
|
|
|
updateParameters(hiddenLayerUpdations, outputLayerUpdation, learning_rate); // subject to change. may want bias to have this matrix too.
|
|
y_hat = modelSetTest(inputMiniBatches[i]);
|
|
|
|
if (UI) {
|
|
MLPPANNOld::UI(epoch, cost_prev, y_hat, outputMiniBatches[i]);
|
|
}
|
|
}
|
|
epoch++;
|
|
if (epoch > max_epoch) {
|
|
break;
|
|
}
|
|
}
|
|
forwardPass();
|
|
}
|
|
|
|
void MLPPANNOld::Adam(real_t learning_rate, int max_epoch, int mini_batch_size, real_t b1, real_t b2, real_t e, bool UI) {
|
|
class MLPPCost cost;
|
|
MLPPLinAlgOld 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(inputSet, outputSet, n_mini_batch);
|
|
auto inputMiniBatches = std::get<0>(batches);
|
|
auto outputMiniBatches = 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 = applyLearningRateScheduler(initial_learning_rate, decayConstant, epoch, dropRate);
|
|
for (int i = 0; i < n_mini_batch; i++) {
|
|
std::vector<real_t> y_hat = modelSetTest(inputMiniBatches[i]);
|
|
cost_prev = Cost(y_hat, outputMiniBatches[i]);
|
|
|
|
auto grads = computeGradients(y_hat, outputMiniBatches[i]);
|
|
auto cumulativeHiddenLayerWGrad = std::get<0>(grads);
|
|
auto outputWGrad = std::get<1>(grads);
|
|
|
|
if (!network.empty() && m_hidden.empty() && v_hidden.empty()) { // Initing our tensor
|
|
m_hidden = alg.resize(m_hidden, cumulativeHiddenLayerWGrad);
|
|
v_hidden = alg.resize(v_hidden, cumulativeHiddenLayerWGrad);
|
|
}
|
|
|
|
if (m_output.empty() && v_output.empty()) {
|
|
m_output.resize(outputWGrad.size());
|
|
v_output.resize(outputWGrad.size());
|
|
}
|
|
|
|
m_hidden = alg.addition(alg.scalarMultiply(b1, m_hidden), alg.scalarMultiply(1 - b1, cumulativeHiddenLayerWGrad));
|
|
v_hidden = alg.addition(alg.scalarMultiply(b2, v_hidden), alg.scalarMultiply(1 - b2, alg.exponentiate(cumulativeHiddenLayerWGrad, 2)));
|
|
|
|
m_output = alg.addition(alg.scalarMultiply(b1, m_output), alg.scalarMultiply(1 - b1, outputWGrad));
|
|
v_output = alg.addition(alg.scalarMultiply(b2, v_output), alg.scalarMultiply(1 - b2, alg.exponentiate(outputWGrad, 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>>> hiddenLayerUpdations = alg.scalarMultiply(learning_rate / n, alg.elementWiseDivision(m_hidden_hat, alg.scalarAdd(e, alg.sqrt(v_hidden_hat))));
|
|
std::vector<real_t> outputLayerUpdation = alg.scalarMultiply(learning_rate / n, alg.elementWiseDivision(m_output_hat, alg.scalarAdd(e, alg.sqrt(v_output_hat))));
|
|
|
|
updateParameters(hiddenLayerUpdations, outputLayerUpdation, learning_rate); // subject to change. may want bias to have this matrix too.
|
|
y_hat = modelSetTest(inputMiniBatches[i]);
|
|
|
|
if (UI) {
|
|
MLPPANNOld::UI(epoch, cost_prev, y_hat, outputMiniBatches[i]);
|
|
}
|
|
}
|
|
epoch++;
|
|
if (epoch > max_epoch) {
|
|
break;
|
|
}
|
|
}
|
|
forwardPass();
|
|
}
|
|
|
|
void MLPPANNOld::Adamax(real_t learning_rate, int max_epoch, int mini_batch_size, real_t b1, real_t b2, real_t e, bool UI) {
|
|
class MLPPCost cost;
|
|
MLPPLinAlgOld 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(inputSet, outputSet, n_mini_batch);
|
|
auto inputMiniBatches = std::get<0>(batches);
|
|
auto outputMiniBatches = 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 = applyLearningRateScheduler(initial_learning_rate, decayConstant, epoch, dropRate);
|
|
for (int i = 0; i < n_mini_batch; i++) {
|
|
std::vector<real_t> y_hat = modelSetTest(inputMiniBatches[i]);
|
|
cost_prev = Cost(y_hat, outputMiniBatches[i]);
|
|
|
|
auto grads = computeGradients(y_hat, outputMiniBatches[i]);
|
|
auto cumulativeHiddenLayerWGrad = std::get<0>(grads);
|
|
auto outputWGrad = std::get<1>(grads);
|
|
|
|
if (!network.empty() && m_hidden.empty() && u_hidden.empty()) { // Initing our tensor
|
|
m_hidden = alg.resize(m_hidden, cumulativeHiddenLayerWGrad);
|
|
u_hidden = alg.resize(u_hidden, cumulativeHiddenLayerWGrad);
|
|
}
|
|
|
|
if (m_output.empty() && u_output.empty()) {
|
|
m_output.resize(outputWGrad.size());
|
|
u_output.resize(outputWGrad.size());
|
|
}
|
|
|
|
m_hidden = alg.addition(alg.scalarMultiply(b1, m_hidden), alg.scalarMultiply(1 - b1, cumulativeHiddenLayerWGrad));
|
|
u_hidden = alg.max(alg.scalarMultiply(b2, u_hidden), alg.abs(cumulativeHiddenLayerWGrad));
|
|
|
|
m_output = alg.addition(alg.scalarMultiply(b1, m_output), alg.scalarMultiply(1 - b1, outputWGrad));
|
|
u_output = alg.max(alg.scalarMultiply(b2, u_output), alg.abs(outputWGrad));
|
|
|
|
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>>> hiddenLayerUpdations = alg.scalarMultiply(learning_rate / n, alg.elementWiseDivision(m_hidden_hat, alg.scalarAdd(e, u_hidden)));
|
|
std::vector<real_t> outputLayerUpdation = alg.scalarMultiply(learning_rate / n, alg.elementWiseDivision(m_output_hat, alg.scalarAdd(e, u_output)));
|
|
|
|
updateParameters(hiddenLayerUpdations, outputLayerUpdation, learning_rate); // subject to change. may want bias to have this matrix too.
|
|
y_hat = modelSetTest(inputMiniBatches[i]);
|
|
|
|
if (UI) {
|
|
MLPPANNOld::UI(epoch, cost_prev, y_hat, outputMiniBatches[i]);
|
|
}
|
|
}
|
|
epoch++;
|
|
if (epoch > max_epoch) {
|
|
break;
|
|
}
|
|
}
|
|
forwardPass();
|
|
}
|
|
|
|
void MLPPANNOld::Nadam(real_t learning_rate, int max_epoch, int mini_batch_size, real_t b1, real_t b2, real_t e, bool UI) {
|
|
class MLPPCost cost;
|
|
MLPPLinAlgOld 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(inputSet, outputSet, n_mini_batch);
|
|
auto inputMiniBatches = std::get<0>(batches);
|
|
auto outputMiniBatches = 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 = applyLearningRateScheduler(initial_learning_rate, decayConstant, epoch, dropRate);
|
|
for (int i = 0; i < n_mini_batch; i++) {
|
|
std::vector<real_t> y_hat = modelSetTest(inputMiniBatches[i]);
|
|
cost_prev = Cost(y_hat, outputMiniBatches[i]);
|
|
|
|
auto grads = computeGradients(y_hat, outputMiniBatches[i]);
|
|
auto cumulativeHiddenLayerWGrad = std::get<0>(grads);
|
|
auto outputWGrad = std::get<1>(grads);
|
|
|
|
if (!network.empty() && m_hidden.empty() && v_hidden.empty()) { // Initing our tensor
|
|
m_hidden = alg.resize(m_hidden, cumulativeHiddenLayerWGrad);
|
|
v_hidden = alg.resize(v_hidden, cumulativeHiddenLayerWGrad);
|
|
}
|
|
|
|
if (m_output.empty() && v_output.empty()) {
|
|
m_output.resize(outputWGrad.size());
|
|
v_output.resize(outputWGrad.size());
|
|
}
|
|
|
|
m_hidden = alg.addition(alg.scalarMultiply(b1, m_hidden), alg.scalarMultiply(1 - b1, cumulativeHiddenLayerWGrad));
|
|
v_hidden = alg.addition(alg.scalarMultiply(b2, v_hidden), alg.scalarMultiply(1 - b2, alg.exponentiate(cumulativeHiddenLayerWGrad, 2)));
|
|
|
|
m_output = alg.addition(alg.scalarMultiply(b1, m_output), alg.scalarMultiply(1 - b1, outputWGrad));
|
|
v_output = alg.addition(alg.scalarMultiply(b2, v_output), alg.scalarMultiply(1 - b2, alg.exponentiate(outputWGrad, 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)), cumulativeHiddenLayerWGrad));
|
|
|
|
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)), outputWGrad));
|
|
|
|
std::vector<std::vector<std::vector<real_t>>> hiddenLayerUpdations = alg.scalarMultiply(learning_rate / n, alg.elementWiseDivision(m_hidden_final, alg.scalarAdd(e, alg.sqrt(v_hidden_hat))));
|
|
std::vector<real_t> outputLayerUpdation = alg.scalarMultiply(learning_rate / n, alg.elementWiseDivision(m_output_final, alg.scalarAdd(e, alg.sqrt(v_output_hat))));
|
|
|
|
updateParameters(hiddenLayerUpdations, outputLayerUpdation, learning_rate); // subject to change. may want bias to have this matrix too.
|
|
y_hat = modelSetTest(inputMiniBatches[i]);
|
|
|
|
if (UI) {
|
|
MLPPANNOld::UI(epoch, cost_prev, y_hat, outputMiniBatches[i]);
|
|
}
|
|
}
|
|
epoch++;
|
|
if (epoch > max_epoch) {
|
|
break;
|
|
}
|
|
}
|
|
forwardPass();
|
|
}
|
|
|
|
void MLPPANNOld::AMSGrad(real_t learning_rate, int max_epoch, int mini_batch_size, real_t b1, real_t b2, real_t e, bool UI) {
|
|
class MLPPCost cost;
|
|
MLPPLinAlgOld 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(inputSet, outputSet, n_mini_batch);
|
|
auto inputMiniBatches = std::get<0>(batches);
|
|
auto outputMiniBatches = 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 = applyLearningRateScheduler(initial_learning_rate, decayConstant, epoch, dropRate);
|
|
for (int i = 0; i < n_mini_batch; i++) {
|
|
std::vector<real_t> y_hat = modelSetTest(inputMiniBatches[i]);
|
|
cost_prev = Cost(y_hat, outputMiniBatches[i]);
|
|
|
|
auto grads = computeGradients(y_hat, outputMiniBatches[i]);
|
|
auto cumulativeHiddenLayerWGrad = std::get<0>(grads);
|
|
auto outputWGrad = std::get<1>(grads);
|
|
|
|
if (!network.empty() && m_hidden.empty() && v_hidden.empty()) { // Initing our tensor
|
|
m_hidden = alg.resize(m_hidden, cumulativeHiddenLayerWGrad);
|
|
v_hidden = alg.resize(v_hidden, cumulativeHiddenLayerWGrad);
|
|
v_hidden_hat = alg.resize(v_hidden_hat, cumulativeHiddenLayerWGrad);
|
|
}
|
|
|
|
if (m_output.empty() && v_output.empty()) {
|
|
m_output.resize(outputWGrad.size());
|
|
v_output.resize(outputWGrad.size());
|
|
v_output_hat.resize(outputWGrad.size());
|
|
}
|
|
|
|
m_hidden = alg.addition(alg.scalarMultiply(b1, m_hidden), alg.scalarMultiply(1 - b1, cumulativeHiddenLayerWGrad));
|
|
v_hidden = alg.addition(alg.scalarMultiply(b2, v_hidden), alg.scalarMultiply(1 - b2, alg.exponentiate(cumulativeHiddenLayerWGrad, 2)));
|
|
|
|
m_output = alg.addition(alg.scalarMultiply(b1, m_output), alg.scalarMultiply(1 - b1, outputWGrad));
|
|
v_output = alg.addition(alg.scalarMultiply(b2, v_output), alg.scalarMultiply(1 - b2, alg.exponentiate(outputWGrad, 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>>> hiddenLayerUpdations = alg.scalarMultiply(learning_rate / n, alg.elementWiseDivision(m_hidden, alg.scalarAdd(e, alg.sqrt(v_hidden_hat))));
|
|
std::vector<real_t> outputLayerUpdation = alg.scalarMultiply(learning_rate / n, alg.elementWiseDivision(m_output, alg.scalarAdd(e, alg.sqrt(v_output_hat))));
|
|
|
|
updateParameters(hiddenLayerUpdations, outputLayerUpdation, learning_rate); // subject to change. may want bias to have this matrix too.
|
|
y_hat = modelSetTest(inputMiniBatches[i]);
|
|
|
|
if (UI) {
|
|
MLPPANNOld::UI(epoch, cost_prev, y_hat, outputMiniBatches[i]);
|
|
}
|
|
}
|
|
epoch++;
|
|
if (epoch > max_epoch) {
|
|
break;
|
|
}
|
|
}
|
|
forwardPass();
|
|
}
|
|
|
|
real_t MLPPANNOld::score() {
|
|
MLPPUtilities util;
|
|
forwardPass();
|
|
return util.performance(y_hat, outputSet);
|
|
}
|
|
|
|
void MLPPANNOld::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, outputLayer->weights, outputLayer->bias, true, network.size() + 1);
|
|
} else {
|
|
util.saveParameters(fileName, outputLayer->weights, outputLayer->bias, false, network.size() + 1);
|
|
}
|
|
}
|
|
|
|
void MLPPANNOld::setLearningRateScheduler(std::string type, real_t decayConstant) {
|
|
lrScheduler = type;
|
|
MLPPANNOld::decayConstant = decayConstant;
|
|
}
|
|
|
|
void MLPPANNOld::setLearningRateScheduler(std::string type, real_t decayConstant, real_t dropRate) {
|
|
lrScheduler = type;
|
|
MLPPANNOld::decayConstant = decayConstant;
|
|
MLPPANNOld::dropRate = dropRate;
|
|
}
|
|
|
|
// https://en.wikipedia.org/wiki/Learning_rate
|
|
// Learning Rate Decay (C2W2L09) - Andrew Ng - Deep Learning Specialization
|
|
real_t MLPPANNOld::applyLearningRateScheduler(real_t learningRate, real_t decayConstant, real_t epoch, real_t dropRate) {
|
|
if (lrScheduler == "Time") {
|
|
return learningRate / (1 + decayConstant * epoch);
|
|
} else if (lrScheduler == "Epoch") {
|
|
return learningRate * (decayConstant / std::sqrt(epoch));
|
|
} else if (lrScheduler == "Step") {
|
|
return learningRate * std::pow(decayConstant, int((1 + epoch) / dropRate)); // Utilizing an explicit int conversion implicitly takes the floor.
|
|
} else if (lrScheduler == "Exponential") {
|
|
return learningRate * std::exp(-decayConstant * epoch);
|
|
}
|
|
return learningRate;
|
|
}
|
|
|
|
void MLPPANNOld::addLayer(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, inputSet, 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 MLPPANNOld::addOutputLayer(std::string activation, std::string loss, std::string weightInit, std::string reg, real_t lambda, real_t alpha) {
|
|
if (!network.empty()) {
|
|
outputLayer = new MLPPOldOutputLayer(network[network.size() - 1].n_hidden, activation, loss, network[network.size() - 1].a, weightInit, reg, lambda, alpha);
|
|
} else {
|
|
outputLayer = new MLPPOldOutputLayer(k, activation, loss, inputSet, weightInit, reg, lambda, alpha);
|
|
}
|
|
}
|
|
|
|
real_t MLPPANNOld::Cost(std::vector<real_t> y_hat, std::vector<real_t> y) {
|
|
MLPPReg regularization;
|
|
class MLPPCost cost;
|
|
real_t totalRegTerm = 0;
|
|
|
|
auto cost_function = outputLayer->cost_map[outputLayer->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 (cost.*cost_function)(y_hat, y) + totalRegTerm + regularization.regTerm(outputLayer->weights, outputLayer->lambda, outputLayer->alpha, outputLayer->reg);
|
|
}
|
|
|
|
void MLPPANNOld::forwardPass() {
|
|
if (!network.empty()) {
|
|
network[0].input = inputSet;
|
|
network[0].forwardPass();
|
|
|
|
for (uint32_t i = 1; i < network.size(); i++) {
|
|
network[i].input = network[i - 1].a;
|
|
network[i].forwardPass();
|
|
}
|
|
outputLayer->input = network[network.size() - 1].a;
|
|
} else {
|
|
outputLayer->input = inputSet;
|
|
}
|
|
outputLayer->forwardPass();
|
|
y_hat = outputLayer->a;
|
|
}
|
|
|
|
void MLPPANNOld::updateParameters(std::vector<std::vector<std::vector<real_t>>> hiddenLayerUpdations, std::vector<real_t> outputLayerUpdation, real_t learning_rate) {
|
|
MLPPLinAlgOld alg;
|
|
|
|
outputLayer->weights = alg.subtraction(outputLayer->weights, outputLayerUpdation);
|
|
outputLayer->bias -= learning_rate * alg.sum_elements(outputLayer->delta) / n;
|
|
|
|
if (!network.empty()) {
|
|
network[network.size() - 1].weights = alg.subtraction(network[network.size() - 1].weights, hiddenLayerUpdations[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, hiddenLayerUpdations[(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>> MLPPANNOld::computeGradients(std::vector<real_t> y_hat, std::vector<real_t> outputSet) {
|
|
// std::cout << "BEGIN" << std::endl;
|
|
class MLPPCost cost;
|
|
MLPPActivationOld avn;
|
|
MLPPLinAlgOld alg;
|
|
MLPPReg regularization;
|
|
|
|
std::vector<std::vector<std::vector<real_t>>> cumulativeHiddenLayerWGrad; // Tensor containing ALL hidden grads.
|
|
|
|
auto costDeriv = outputLayer->costDeriv_map[outputLayer->cost];
|
|
auto outputAvn = outputLayer->activation_map[outputLayer->activation];
|
|
outputLayer->delta = alg.hadamard_product((cost.*costDeriv)(y_hat, outputSet), (avn.*outputAvn)(outputLayer->z, 1));
|
|
std::vector<real_t> outputWGrad = alg.mat_vec_mult(alg.transpose(outputLayer->input), outputLayer->delta);
|
|
outputWGrad = alg.addition(outputWGrad, regularization.regDerivTerm(outputLayer->weights, outputLayer->lambda, outputLayer->alpha, outputLayer->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(outputLayer->delta, outputLayer->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);
|
|
|
|
cumulativeHiddenLayerWGrad.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);
|
|
cumulativeHiddenLayerWGrad.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 { cumulativeHiddenLayerWGrad, outputWGrad };
|
|
}
|
|
|
|
void MLPPANNOld::UI(int epoch, real_t cost_prev, std::vector<real_t> y_hat, std::vector<real_t> outputSet) {
|
|
MLPPUtilities::CostInfo(epoch, cost_prev, Cost(y_hat, outputSet));
|
|
std::cout << "Layer " << network.size() + 1 << ": " << std::endl;
|
|
MLPPUtilities::UI(outputLayer->weights, outputLayer->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);
|
|
}
|
|
}
|
|
}
|