Relintai/pmlpp
1
0
Fork 0
mirror of https://github.com/Relintai/pmlpp.git synced 2024-11-14 14:07:18 +01:00
Code Issues Packages Projects Releases Wiki Activity
43e1b8d1fc
pmlpp/mlpp/ann/ann.cpp

763 lines
30 KiB
C++
Raw Normal View History

Added https://github.com/novak-99/MLPP as a base, without the included datasets.
2023-01-23 21:13:26 +01:00
//
// ANN.cpp
//
// Created by Marc Melikyan on 11/4/20.
//
Include cleanups.
2023-01-24 18:12:23 +01:00
#include "ann.h"
#include "../activation/activation.h"
Clang format.
2023-01-24 19:00:54 +01:00
#include "../cost/cost.h"
Include cleanups.
2023-01-24 18:12:23 +01:00
#include "../lin_alg/lin_alg.h"
#include "../regularization/reg.h"
#include "../utilities/utilities.h"
Added https://github.com/novak-99/MLPP as a base, without the included datasets.
2023-01-23 21:13:26 +01:00
#include <cmath>
Clang format.
2023-01-24 19:00:54 +01:00
#include <iostream>
Added https://github.com/novak-99/MLPP as a base, without the included datasets.
2023-01-23 21:13:26 +01:00
#include <random>
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
MLPPANN::MLPPANN(std::vector<std::vector<double>> inputSet, std::vector<double> outputSet) :
Clang format.
2023-01-24 19:00:54 +01:00
inputSet(inputSet), outputSet(outputSet), n(inputSet.size()), k(inputSet[0].size()), lrScheduler("None"), decayConstant(0), dropRate(0) {
}
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
MLPPANN::~MLPPANN() {
Clang format.
2023-01-24 19:00:54 +01:00
delete outputLayer;
}
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
std::vector<double> MLPPANN::modelSetTest(std::vector<std::vector<double>> X) {
Clang format.
2023-01-24 19:00:54 +01:00
if (!network.empty()) {
network[0].input = X;
network[0].forwardPass();
for (int 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;
}
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
double MLPPANN::modelTest(std::vector<double> x) {
Clang format.
2023-01-24 19:00:54 +01:00
if (!network.empty()) {
network[0].Test(x);
for (int 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;
}
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
void MLPPANN::gradientDescent(double learning_rate, int max_epoch, bool UI) {
Prefix Cost with MLPP.
2023-01-24 19:37:08 +01:00
class MLPPCost cost;
Prefixed LinAlg with MLPP.
2023-01-25 00:29:02 +01:00
MLPPLinAlg alg;
Clang format.
2023-01-24 19:00:54 +01:00
double cost_prev = 0;
int epoch = 1;
forwardPass();
double 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 [cumulativeHiddenLayerWGrad, outputWGrad] = computeGradients(y_hat, outputSet);
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) {
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
MLPPANN::UI(epoch, cost_prev, y_hat, outputSet);
Clang format.
2023-01-24 19:00:54 +01:00
}
epoch++;
if (epoch > max_epoch) {
break;
}
}
}
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
void MLPPANN::SGD(double learning_rate, int max_epoch, bool UI) {
Prefix Cost with MLPP.
2023-01-24 19:37:08 +01:00
class MLPPCost cost;
Prefixed LinAlg with MLPP.
2023-01-25 00:29:02 +01:00
MLPPLinAlg alg;
Clang format.
2023-01-24 19:00:54 +01:00
double cost_prev = 0;
int epoch = 1;
double 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<double> y_hat = modelSetTest({ inputSet[outputIndex] });
cost_prev = Cost({ y_hat }, { outputSet[outputIndex] });
auto [cumulativeHiddenLayerWGrad, outputWGrad] = computeGradients(y_hat, { outputSet[outputIndex] });
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) {
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
MLPPANN::UI(epoch, cost_prev, y_hat, { outputSet[outputIndex] });
Clang format.
2023-01-24 19:00:54 +01:00
}
epoch++;
if (epoch > max_epoch) {
break;
}
}
forwardPass();
}
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
void MLPPANN::MBGD(double learning_rate, int max_epoch, int mini_batch_size, bool UI) {
Prefix Cost with MLPP.
2023-01-24 19:37:08 +01:00
class MLPPCost cost;
Prefixed LinAlg with MLPP.
2023-01-25 00:29:02 +01:00
MLPPLinAlg alg;
Clang format.
2023-01-24 19:00:54 +01:00
double cost_prev = 0;
int epoch = 1;
double 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 [inputMiniBatches, outputMiniBatches] = Utilities::createMiniBatches(inputSet, outputSet, n_mini_batch);
while (true) {
learning_rate = applyLearningRateScheduler(initial_learning_rate, decayConstant, epoch, dropRate);
for (int i = 0; i < n_mini_batch; i++) {
std::vector<double> y_hat = modelSetTest(inputMiniBatches[i]);
cost_prev = Cost(y_hat, outputMiniBatches[i]);
auto [cumulativeHiddenLayerWGrad, outputWGrad] = computeGradients(y_hat, outputMiniBatches[i]);
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) {
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
MLPPANN::UI(epoch, cost_prev, y_hat, outputMiniBatches[i]);
Clang format.
2023-01-24 19:00:54 +01:00
}
}
epoch++;
if (epoch > max_epoch) {
break;
}
}
forwardPass();
}
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
void MLPPANN::Momentum(double learning_rate, int max_epoch, int mini_batch_size, double gamma, bool NAG, bool UI) {
Prefix Cost with MLPP.
2023-01-24 19:37:08 +01:00
class MLPPCost cost;
Prefixed LinAlg with MLPP.
2023-01-25 00:29:02 +01:00
MLPPLinAlg alg;
Clang format.
2023-01-24 19:00:54 +01:00
double cost_prev = 0;
int epoch = 1;
double 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 [inputMiniBatches, outputMiniBatches] = Utilities::createMiniBatches(inputSet, outputSet, n_mini_batch);
// Initializing necessary components for Adam.
std::vector<std::vector<std::vector<double>>> v_hidden;
std::vector<double> v_output;
while (true) {
learning_rate = applyLearningRateScheduler(initial_learning_rate, decayConstant, epoch, dropRate);
for (int i = 0; i < n_mini_batch; i++) {
std::vector<double> y_hat = modelSetTest(inputMiniBatches[i]);
cost_prev = Cost(y_hat, outputMiniBatches[i]);
auto [cumulativeHiddenLayerWGrad, outputWGrad] = computeGradients(y_hat, outputMiniBatches[i]);
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) {
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
MLPPANN::UI(epoch, cost_prev, y_hat, outputMiniBatches[i]);
Clang format.
2023-01-24 19:00:54 +01:00
}
}
epoch++;
if (epoch > max_epoch) {
break;
}
}
forwardPass();
}
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
void MLPPANN::Adagrad(double learning_rate, int max_epoch, int mini_batch_size, double e, bool UI) {
Prefix Cost with MLPP.
2023-01-24 19:37:08 +01:00
class MLPPCost cost;
Prefixed LinAlg with MLPP.
2023-01-25 00:29:02 +01:00
MLPPLinAlg alg;
Clang format.
2023-01-24 19:00:54 +01:00
double cost_prev = 0;
int epoch = 1;
double 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 [inputMiniBatches, outputMiniBatches] = Utilities::createMiniBatches(inputSet, outputSet, n_mini_batch);
// Initializing necessary components for Adam.
std::vector<std::vector<std::vector<double>>> v_hidden;
std::vector<double> v_output;
while (true) {
learning_rate = applyLearningRateScheduler(initial_learning_rate, decayConstant, epoch, dropRate);
for (int i = 0; i < n_mini_batch; i++) {
std::vector<double> y_hat = modelSetTest(inputMiniBatches[i]);
cost_prev = Cost(y_hat, outputMiniBatches[i]);
auto [cumulativeHiddenLayerWGrad, outputWGrad] = computeGradients(y_hat, outputMiniBatches[i]);
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<double>>> hiddenLayerUpdations = alg.scalarMultiply(learning_rate / n, alg.elementWiseDivision(cumulativeHiddenLayerWGrad, alg.scalarAdd(e, alg.sqrt(v_hidden))));
std::vector<double> 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) {
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
MLPPANN::UI(epoch, cost_prev, y_hat, outputMiniBatches[i]);
Clang format.
2023-01-24 19:00:54 +01:00
}
}
epoch++;
if (epoch > max_epoch) {
break;
}
}
forwardPass();
}
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
void MLPPANN::Adadelta(double learning_rate, int max_epoch, int mini_batch_size, double b1, double e, bool UI) {
Prefix Cost with MLPP.
2023-01-24 19:37:08 +01:00
class MLPPCost cost;
Prefixed LinAlg with MLPP.
2023-01-25 00:29:02 +01:00
MLPPLinAlg alg;
Clang format.
2023-01-24 19:00:54 +01:00
double cost_prev = 0;
int epoch = 1;
double 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 [inputMiniBatches, outputMiniBatches] = Utilities::createMiniBatches(inputSet, outputSet, n_mini_batch);
// Initializing necessary components for Adam.
std::vector<std::vector<std::vector<double>>> v_hidden;
std::vector<double> v_output;
while (true) {
learning_rate = applyLearningRateScheduler(initial_learning_rate, decayConstant, epoch, dropRate);
for (int i = 0; i < n_mini_batch; i++) {
std::vector<double> y_hat = modelSetTest(inputMiniBatches[i]);
cost_prev = Cost(y_hat, outputMiniBatches[i]);
auto [cumulativeHiddenLayerWGrad, outputWGrad] = computeGradients(y_hat, outputMiniBatches[i]);
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<double>>> hiddenLayerUpdations = alg.scalarMultiply(learning_rate / n, alg.elementWiseDivision(cumulativeHiddenLayerWGrad, alg.scalarAdd(e, alg.sqrt(v_hidden))));
std::vector<double> 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) {
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
MLPPANN::UI(epoch, cost_prev, y_hat, outputMiniBatches[i]);
Clang format.
2023-01-24 19:00:54 +01:00
}
}
epoch++;
if (epoch > max_epoch) {
break;
}
}
forwardPass();
}
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
void MLPPANN::Adam(double learning_rate, int max_epoch, int mini_batch_size, double b1, double b2, double e, bool UI) {
Prefix Cost with MLPP.
2023-01-24 19:37:08 +01:00
class MLPPCost cost;
Prefixed LinAlg with MLPP.
2023-01-25 00:29:02 +01:00
MLPPLinAlg alg;
Clang format.
2023-01-24 19:00:54 +01:00
double cost_prev = 0;
int epoch = 1;
double 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 [inputMiniBatches, outputMiniBatches] = Utilities::createMiniBatches(inputSet, outputSet, n_mini_batch);
// Initializing necessary components for Adam.
std::vector<std::vector<std::vector<double>>> m_hidden;
std::vector<std::vector<std::vector<double>>> v_hidden;
std::vector<double> m_output;
std::vector<double> v_output;
while (true) {
learning_rate = applyLearningRateScheduler(initial_learning_rate, decayConstant, epoch, dropRate);
for (int i = 0; i < n_mini_batch; i++) {
std::vector<double> y_hat = modelSetTest(inputMiniBatches[i]);
cost_prev = Cost(y_hat, outputMiniBatches[i]);
auto [cumulativeHiddenLayerWGrad, outputWGrad] = computeGradients(y_hat, outputMiniBatches[i]);
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<double>>> m_hidden_hat = alg.scalarMultiply(1 / (1 - std::pow(b1, epoch)), m_hidden);
std::vector<std::vector<std::vector<double>>> v_hidden_hat = alg.scalarMultiply(1 / (1 - std::pow(b2, epoch)), v_hidden);
std::vector<double> m_output_hat = alg.scalarMultiply(1 / (1 - std::pow(b1, epoch)), m_output);
std::vector<double> v_output_hat = alg.scalarMultiply(1 / (1 - std::pow(b2, epoch)), v_output);
std::vector<std::vector<std::vector<double>>> hiddenLayerUpdations = alg.scalarMultiply(learning_rate / n, alg.elementWiseDivision(m_hidden_hat, alg.scalarAdd(e, alg.sqrt(v_hidden_hat))));
std::vector<double> 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) {
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
MLPPANN::UI(epoch, cost_prev, y_hat, outputMiniBatches[i]);
Clang format.
2023-01-24 19:00:54 +01:00
}
}
epoch++;
if (epoch > max_epoch) {
break;
}
}
forwardPass();
}
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
void MLPPANN::Adamax(double learning_rate, int max_epoch, int mini_batch_size, double b1, double b2, double e, bool UI) {
Prefix Cost with MLPP.
2023-01-24 19:37:08 +01:00
class MLPPCost cost;
Prefixed LinAlg with MLPP.
2023-01-25 00:29:02 +01:00
MLPPLinAlg alg;
Clang format.
2023-01-24 19:00:54 +01:00
double cost_prev = 0;
int epoch = 1;
double 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 [inputMiniBatches, outputMiniBatches] = Utilities::createMiniBatches(inputSet, outputSet, n_mini_batch);
// Initializing necessary components for Adam.
std::vector<std::vector<std::vector<double>>> m_hidden;
std::vector<std::vector<std::vector<double>>> u_hidden;
std::vector<double> m_output;
std::vector<double> u_output;
while (true) {
learning_rate = applyLearningRateScheduler(initial_learning_rate, decayConstant, epoch, dropRate);
for (int i = 0; i < n_mini_batch; i++) {
std::vector<double> y_hat = modelSetTest(inputMiniBatches[i]);
cost_prev = Cost(y_hat, outputMiniBatches[i]);
auto [cumulativeHiddenLayerWGrad, outputWGrad] = computeGradients(y_hat, outputMiniBatches[i]);
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<double>>> m_hidden_hat = alg.scalarMultiply(1 / (1 - std::pow(b1, epoch)), m_hidden);
std::vector<double> m_output_hat = alg.scalarMultiply(1 / (1 - std::pow(b1, epoch)), m_output);
std::vector<std::vector<std::vector<double>>> hiddenLayerUpdations = alg.scalarMultiply(learning_rate / n, alg.elementWiseDivision(m_hidden_hat, alg.scalarAdd(e, u_hidden)));
std::vector<double> 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) {
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
MLPPANN::UI(epoch, cost_prev, y_hat, outputMiniBatches[i]);
Clang format.
2023-01-24 19:00:54 +01:00
}
}
epoch++;
if (epoch > max_epoch) {
break;
}
}
forwardPass();
}
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
void MLPPANN::Nadam(double learning_rate, int max_epoch, int mini_batch_size, double b1, double b2, double e, bool UI) {
Prefix Cost with MLPP.
2023-01-24 19:37:08 +01:00
class MLPPCost cost;
Prefixed LinAlg with MLPP.
2023-01-25 00:29:02 +01:00
MLPPLinAlg alg;
Clang format.
2023-01-24 19:00:54 +01:00
double cost_prev = 0;
int epoch = 1;
double 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 [inputMiniBatches, outputMiniBatches] = Utilities::createMiniBatches(inputSet, outputSet, n_mini_batch);
// Initializing necessary components for Adam.
std::vector<std::vector<std::vector<double>>> m_hidden;
std::vector<std::vector<std::vector<double>>> v_hidden;
std::vector<std::vector<std::vector<double>>> m_hidden_final;
std::vector<double> m_output;
std::vector<double> v_output;
while (true) {
learning_rate = applyLearningRateScheduler(initial_learning_rate, decayConstant, epoch, dropRate);
for (int i = 0; i < n_mini_batch; i++) {
std::vector<double> y_hat = modelSetTest(inputMiniBatches[i]);
cost_prev = Cost(y_hat, outputMiniBatches[i]);
auto [cumulativeHiddenLayerWGrad, outputWGrad] = computeGradients(y_hat, outputMiniBatches[i]);
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<double>>> m_hidden_hat = alg.scalarMultiply(1 / (1 - std::pow(b1, epoch)), m_hidden);
std::vector<std::vector<std::vector<double>>> v_hidden_hat = alg.scalarMultiply(1 / (1 - std::pow(b2, epoch)), v_hidden);
std::vector<std::vector<std::vector<double>>> m_hidden_final = alg.addition(alg.scalarMultiply(b1, m_hidden_hat), alg.scalarMultiply((1 - b1) / (1 - std::pow(b1, epoch)), cumulativeHiddenLayerWGrad));
std::vector<double> m_output_hat = alg.scalarMultiply(1 / (1 - std::pow(b1, epoch)), m_output);
std::vector<double> v_output_hat = alg.scalarMultiply(1 / (1 - std::pow(b2, epoch)), v_output);
std::vector<double> 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<double>>> hiddenLayerUpdations = alg.scalarMultiply(learning_rate / n, alg.elementWiseDivision(m_hidden_final, alg.scalarAdd(e, alg.sqrt(v_hidden_hat))));
std::vector<double> 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) {
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
MLPPANN::UI(epoch, cost_prev, y_hat, outputMiniBatches[i]);
Clang format.
2023-01-24 19:00:54 +01:00
}
}
epoch++;
if (epoch > max_epoch) {
break;
}
}
forwardPass();
}
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
void MLPPANN::AMSGrad(double learning_rate, int max_epoch, int mini_batch_size, double b1, double b2, double e, bool UI) {
Prefix Cost with MLPP.
2023-01-24 19:37:08 +01:00
class MLPPCost cost;
Prefixed LinAlg with MLPP.
2023-01-25 00:29:02 +01:00
MLPPLinAlg alg;
Clang format.
2023-01-24 19:00:54 +01:00
double cost_prev = 0;
int epoch = 1;
double 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 [inputMiniBatches, outputMiniBatches] = Utilities::createMiniBatches(inputSet, outputSet, n_mini_batch);
// Initializing necessary components for Adam.
std::vector<std::vector<std::vector<double>>> m_hidden;
std::vector<std::vector<std::vector<double>>> v_hidden;
std::vector<std::vector<std::vector<double>>> v_hidden_hat;
std::vector<double> m_output;
std::vector<double> v_output;
std::vector<double> 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<double> y_hat = modelSetTest(inputMiniBatches[i]);
cost_prev = Cost(y_hat, outputMiniBatches[i]);
auto [cumulativeHiddenLayerWGrad, outputWGrad] = computeGradients(y_hat, outputMiniBatches[i]);
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<double>>> hiddenLayerUpdations = alg.scalarMultiply(learning_rate / n, alg.elementWiseDivision(m_hidden, alg.scalarAdd(e, alg.sqrt(v_hidden_hat))));
std::vector<double> 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) {
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
MLPPANN::UI(epoch, cost_prev, y_hat, outputMiniBatches[i]);
Clang format.
2023-01-24 19:00:54 +01:00
}
}
epoch++;
if (epoch > max_epoch) {
break;
}
}
forwardPass();
}
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
double MLPPANN::score() {
Clang format.
2023-01-24 19:00:54 +01:00
Utilities util;
forwardPass();
return util.performance(y_hat, outputSet);
}
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
void MLPPANN::save(std::string fileName) {
Clang format.
2023-01-24 19:00:54 +01:00
Utilities util;
if (!network.empty()) {
util.saveParameters(fileName, network[0].weights, network[0].bias, 0, 1);
for (int i = 1; i < network.size(); i++) {
util.saveParameters(fileName, network[i].weights, network[i].bias, 1, i + 1);
}
util.saveParameters(fileName, outputLayer->weights, outputLayer->bias, 1, network.size() + 1);
} else {
util.saveParameters(fileName, outputLayer->weights, outputLayer->bias, 0, network.size() + 1);
}
}
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
void MLPPANN::setLearningRateScheduler(std::string type, double decayConstant) {
Clang format.
2023-01-24 19:00:54 +01:00
lrScheduler = type;
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
MLPPANN::decayConstant = decayConstant;
Clang format.
2023-01-24 19:00:54 +01:00
}
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
void MLPPANN::setLearningRateScheduler(std::string type, double decayConstant, double dropRate) {
Clang format.
2023-01-24 19:00:54 +01:00
lrScheduler = type;
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
MLPPANN::decayConstant = decayConstant;
MLPPANN::dropRate = dropRate;
Clang format.
2023-01-24 19:00:54 +01:00
}
// https://en.wikipedia.org/wiki/Learning_rate
// Learning Rate Decay (C2W2L09) - Andrew Ng - Deep Learning Specialization
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
double MLPPANN::applyLearningRateScheduler(double learningRate, double decayConstant, double epoch, double dropRate) {
Clang format.
2023-01-24 19:00:54 +01:00
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;
}
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
void MLPPANN::addLayer(int n_hidden, std::string activation, std::string weightInit, std::string reg, double lambda, double alpha) {
Clang format.
2023-01-24 19:00:54 +01:00
if (network.empty()) {
Prefixed more classes with MLPP.
2023-01-25 00:21:31 +01:00
network.push_back(MLPPHiddenLayer(n_hidden, activation, inputSet, weightInit, reg, lambda, alpha));
Clang format.
2023-01-24 19:00:54 +01:00
network[0].forwardPass();
} else {
Prefixed more classes with MLPP.
2023-01-25 00:21:31 +01:00
network.push_back(MLPPHiddenLayer(n_hidden, activation, network[network.size() - 1].a, weightInit, reg, lambda, alpha));
Clang format.
2023-01-24 19:00:54 +01:00
network[network.size() - 1].forwardPass();
}
}
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
void MLPPANN::addOutputLayer(std::string activation, std::string loss, std::string weightInit, std::string reg, double lambda, double alpha) {
Prefixed LinAlg with MLPP.
2023-01-25 00:29:02 +01:00
MLPPLinAlg alg;
Clang format.
2023-01-24 19:00:54 +01:00
if (!network.empty()) {
Prefixed more classes with MLPP.
2023-01-25 00:54:50 +01:00
outputLayer = new MLPPOutputLayer(network[network.size() - 1].n_hidden, activation, loss, network[network.size() - 1].a, weightInit, reg, lambda, alpha);
Clang format.
2023-01-24 19:00:54 +01:00
} else {
Prefixed more classes with MLPP.
2023-01-25 00:54:50 +01:00
outputLayer = new MLPPOutputLayer(k, activation, loss, inputSet, weightInit, reg, lambda, alpha);
Clang format.
2023-01-24 19:00:54 +01:00
}
}
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
double MLPPANN::Cost(std::vector<double> y_hat, std::vector<double> y) {
Prefixed more classes with MLPP.
2023-01-25 00:54:50 +01:00
MLPPReg regularization;
Prefix Cost with MLPP.
2023-01-24 19:37:08 +01:00
class MLPPCost cost;
Clang format.
2023-01-24 19:00:54 +01:00
double totalRegTerm = 0;
auto cost_function = outputLayer->cost_map[outputLayer->cost];
if (!network.empty()) {
for (int 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);
}
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
void MLPPANN::forwardPass() {
Clang format.
2023-01-24 19:00:54 +01:00
if (!network.empty()) {
network[0].input = inputSet;
network[0].forwardPass();
for (int 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;
}
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
void MLPPANN::updateParameters(std::vector<std::vector<std::vector<double>>> hiddenLayerUpdations, std::vector<double> outputLayerUpdation, double learning_rate) {
Prefixed LinAlg with MLPP.
2023-01-25 00:29:02 +01:00
MLPPLinAlg alg;
Clang format.
2023-01-24 19:00:54 +01:00
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));
}
}
}
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
std::tuple<std::vector<std::vector<std::vector<double>>>, std::vector<double>> MLPPANN::computeGradients(std::vector<double> y_hat, std::vector<double> outputSet) {
Clang format.
2023-01-24 19:00:54 +01:00
// std::cout << "BEGIN" << std::endl;
Prefix Cost with MLPP.
2023-01-24 19:37:08 +01:00
class MLPPCost cost;
Renamed Activation to MLPPActivation.
2023-01-24 19:23:30 +01:00
MLPPActivation avn;
Prefixed LinAlg with MLPP.
2023-01-25 00:29:02 +01:00
MLPPLinAlg alg;
Prefixed more classes with MLPP.
2023-01-25 00:54:50 +01:00
MLPPReg regularization;
Clang format.
2023-01-24 19:00:54 +01:00
std::vector<std::vector<std::vector<double>>> 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<double> 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<double>> 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--) {
auto 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));
std::vector<std::vector<double>> 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 };
}
Prefix classes with MLPP.
2023-01-24 19:29:29 +01:00
void MLPPANN::UI(int epoch, double cost_prev, std::vector<double> y_hat, std::vector<double> outputSet) {
Clang format.
2023-01-24 19:00:54 +01:00
Utilities::CostInfo(epoch, cost_prev, Cost(y_hat, outputSet));
std::cout << "Layer " << network.size() + 1 << ": " << std::endl;
Utilities::UI(outputLayer->weights, outputLayer->bias);
if (!network.empty()) {
for (int i = network.size() - 1; i >= 0; i--) {
std::cout << "Layer " << i + 1 << ": " << std::endl;
Utilities::UI(network[i].weights, network[i].bias);
}
}
}
Reference in New Issue Copy Permalink