Added old versions for all remaining classes.

SCsub

@@ -63,6 +63,18 @@ sources = [
     "mlpp/softmax_reg/softmax_reg_old.cpp",
     "mlpp/auto_encoder/auto_encoder_old.cpp",
     "mlpp/tanh_reg/tanh_reg_old.cpp",
+    "mlpp/softmax_net/softmax_net_old.cpp",
+    "mlpp/multinomial_nb/multinomial_nb_old.cpp",
+    "mlpp/mann/mann_old.cpp",
+    "mlpp/log_reg/log_reg_old.cpp",
+    "mlpp/lin_reg/lin_reg_old.cpp",
+    "mlpp/gaussian_nb/gaussian_nb_old.cpp",
+    "mlpp/gan/gan_old.cpp",
+    "mlpp/exp_reg/exp_reg_old.cpp",
+    "mlpp/dual_svc/dual_svc_old.cpp",
+    "mlpp/c_log_log_reg/c_log_log_reg_old.cpp",
+    "mlpp/bernoulli_nb/bernoulli_nb_old.cpp",
+    "mlpp/ann/ann_old.cpp",
 
     "test/mlpp_tests.cpp",
 ]

mlpp/ann/ann_old.cpp

@@ -0,0 +1,818 @@
+//
+//  ANN.cpp
+//
+//  Created by Marc Melikyan on 11/4/20.
+//
+
+#include "ann_old.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>
+
+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;
+	MLPPLinAlg 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;
+	MLPPLinAlg 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;
+	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(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;
+	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(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;
+	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(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;
+	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(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;
+	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(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;
+	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(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;
+	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(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;
+	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(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) {
+	MLPPLinAlg 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;
+	MLPPActivation avn;
+	MLPPLinAlg 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);
+		}
+	}
+}

mlpp/ann/ann_old.h

@@ -0,0 +1,73 @@
+#ifndef MLPP_ANN_OLD_H
+#define MLPP_ANN_OLD_H
+
+//
+//  ANN.hpp
+//
+//  Created by Marc Melikyan on 11/4/20.
+//
+
+#include "core/math/math_defs.h"
+
+#include "../hidden_layer/hidden_layer.h"
+#include "../output_layer/output_layer.h"
+
+#include "../hidden_layer/hidden_layer_old.h"
+#include "../output_layer/output_layer_old.h"
+
+#include <string>
+#include <tuple>
+#include <vector>
+
+class MLPPANNOld {
+public:
+	MLPPANNOld(std::vector<std::vector<real_t>> inputSet, std::vector<real_t> outputSet);
+	~MLPPANNOld();
+	std::vector<real_t> modelSetTest(std::vector<std::vector<real_t>> X);
+	real_t modelTest(std::vector<real_t> x);
+	void gradientDescent(real_t learning_rate, int max_epoch, bool UI = false);
+	void SGD(real_t learning_rate, int max_epoch, bool UI = false);
+	void MBGD(real_t learning_rate, int max_epoch, int mini_batch_size, bool UI = false);
+	void Momentum(real_t learning_rate, int max_epoch, int mini_batch_size, real_t gamma, bool NAG, bool UI = false);
+	void Adagrad(real_t learning_rate, int max_epoch, int mini_batch_size, real_t e, bool UI = false);
+	void Adadelta(real_t learning_rate, int max_epoch, int mini_batch_size, real_t b1, real_t e, bool UI = false);
+	void Adam(real_t learning_rate, int max_epoch, int mini_batch_size, real_t b1, real_t b2, real_t e, bool UI = false);
+	void Adamax(real_t learning_rate, int max_epoch, int mini_batch_size, real_t b1, real_t b2, real_t e, bool UI = false);
+	void Nadam(real_t learning_rate, int max_epoch, int mini_batch_size, real_t b1, real_t b2, real_t e, bool UI = false);
+	void AMSGrad(real_t learning_rate, int max_epoch, int mini_batch_size, real_t b1, real_t b2, real_t e, bool UI = false);
+	real_t score();
+	void save(std::string fileName);
+
+	void setLearningRateScheduler(std::string type, real_t decayConstant);
+	void setLearningRateScheduler(std::string type, real_t decayConstant, real_t dropRate);
+
+	void addLayer(int n_hidden, std::string activation, std::string weightInit = "Default", std::string reg = "None", real_t lambda = 0.5, real_t alpha = 0.5);
+	void addOutputLayer(std::string activation, std::string loss, std::string weightInit = "Default", std::string reg = "None", real_t lambda = 0.5, real_t alpha = 0.5);
+
+private:
+	real_t applyLearningRateScheduler(real_t learningRate, real_t decayConstant, real_t epoch, real_t dropRate);
+
+	real_t Cost(std::vector<real_t> y_hat, std::vector<real_t> y);
+
+	void forwardPass();
+	void updateParameters(std::vector<std::vector<std::vector<real_t>>> hiddenLayerUpdations, std::vector<real_t> outputLayerUpdation, real_t learning_rate);
+	std::tuple<std::vector<std::vector<std::vector<real_t>>>, std::vector<real_t>> computeGradients(std::vector<real_t> y_hat, std::vector<real_t> outputSet);
+
+	void UI(int epoch, real_t cost_prev, std::vector<real_t> y_hat, std::vector<real_t> outputSet);
+
+	std::vector<std::vector<real_t>> inputSet;
+	std::vector<real_t> outputSet;
+	std::vector<real_t> y_hat;
+
+	std::vector<MLPPOldHiddenLayer> network;
+	MLPPOldOutputLayer *outputLayer;
+
+	int n;
+	int k;
+
+	std::string lrScheduler;
+	real_t decayConstant;
+	real_t dropRate;
+};
+
+#endif /* ANN_hpp */

mlpp/bernoulli_nb/bernoulli_nb_old.cpp

@@ -0,0 +1,179 @@
+//
+//  BernoulliNB.cpp
+//
+//  Created by Marc Melikyan on 1/17/21.
+//
+
+#include "bernoulli_nb_old.h"
+#include "../data/data.h"
+#include "../lin_alg/lin_alg.h"
+#include "../utilities/utilities.h"
+
+#include <iostream>
+#include <random>
+
+MLPPBernoulliNBOld::MLPPBernoulliNBOld(std::vector<std::vector<real_t>> p_inputSet, std::vector<real_t> p_outputSet) {
+	inputSet = p_inputSet;
+	outputSet = p_outputSet;
+	class_num = 2;
+
+	y_hat.resize(outputSet.size());
+	Evaluate();
+}
+
+std::vector<real_t> MLPPBernoulliNBOld::modelSetTest(std::vector<std::vector<real_t>> X) {
+	std::vector<real_t> y_hat;
+	for (uint32_t i = 0; i < X.size(); i++) {
+		y_hat.push_back(modelTest(X[i]));
+	}
+	return y_hat;
+}
+
+real_t MLPPBernoulliNBOld::modelTest(std::vector<real_t> x) {
+	real_t score_0 = 1;
+	real_t score_1 = 1;
+
+	std::vector<int> foundIndices;
+
+	for (uint32_t j = 0; j < x.size(); j++) {
+		for (uint32_t k = 0; k < vocab.size(); k++) {
+			if (x[j] == vocab[k]) {
+				score_0 *= theta[0][vocab[k]];
+				score_1 *= theta[1][vocab[k]];
+
+				foundIndices.push_back(k);
+			}
+		}
+	}
+
+	for (uint32_t i = 0; i < vocab.size(); i++) {
+		bool found = false;
+		for (uint32_t j = 0; j < foundIndices.size(); j++) {
+			if (vocab[i] == vocab[foundIndices[j]]) {
+				found = true;
+			}
+		}
+		if (!found) {
+			score_0 *= 1 - theta[0][vocab[i]];
+			score_1 *= 1 - theta[1][vocab[i]];
+		}
+	}
+
+	score_0 *= prior_0;
+	score_1 *= prior_1;
+
+	// Assigning the traning example to a class
+
+	if (score_0 > score_1) {
+		return 0;
+	} else {
+		return 1;
+	}
+}
+
+real_t MLPPBernoulliNBOld::score() {
+	MLPPUtilities util;
+	return util.performance(y_hat, outputSet);
+}
+
+void MLPPBernoulliNBOld::computeVocab() {
+	MLPPLinAlg alg;
+	MLPPData data;
+	vocab = data.vecToSet<real_t>(alg.flatten(inputSet));
+}
+
+void MLPPBernoulliNBOld::computeTheta() {
+	// Resizing theta for the sake of ease & proper access of the elements.
+	theta.resize(class_num);
+
+	// Setting all values in the hasmap by default to 0.
+	for (int i = class_num - 1; i >= 0; i--) {
+		for (uint32_t j = 0; j < vocab.size(); j++) {
+			theta[i][vocab[j]] = 0;
+		}
+	}
+
+	for (uint32_t i = 0; i < inputSet.size(); i++) {
+		for (uint32_t j = 0; j < inputSet[0].size(); j++) {
+			theta[outputSet[i]][inputSet[i][j]]++;
+		}
+	}
+
+	for (uint32_t i = 0; i < theta.size(); i++) {
+		for (uint32_t j = 0; j < theta[i].size(); j++) {
+			if (i == 0) {
+				theta[i][j] /= prior_0 * y_hat.size();
+			} else {
+				theta[i][j] /= prior_1 * y_hat.size();
+			}
+		}
+	}
+}
+
+void MLPPBernoulliNBOld::Evaluate() {
+	for (uint32_t i = 0; i < outputSet.size(); i++) {
+		// Pr(B | A) * Pr(A)
+		real_t score_0 = 1;
+		real_t score_1 = 1;
+
+		real_t sum = 0;
+		for (uint32_t ii = 0; ii < outputSet.size(); ii++) {
+			if (outputSet[ii] == 1) {
+				sum += outputSet[ii];
+			}
+		}
+
+		// Easy computation of priors, i.e. Pr(C_k)
+		prior_1 = sum / y_hat.size();
+		prior_0 = 1 - prior_1;
+
+		// Evaluating Theta...
+		computeTheta();
+
+		// Evaluating the vocab set...
+		computeVocab();
+
+		std::vector<int> foundIndices;
+
+		for (uint32_t j = 0; j < inputSet.size(); j++) {
+			for (uint32_t k = 0; k < vocab.size(); k++) {
+				if (inputSet[i][j] == vocab[k]) {
+					score_0 += std::log(theta[0][vocab[k]]);
+					score_1 += std::log(theta[1][vocab[k]]);
+
+					foundIndices.push_back(k);
+				}
+			}
+		}
+
+		for (uint32_t ii = 0; ii < vocab.size(); ii++) {
+			bool found = false;
+			for (uint32_t j = 0; j < foundIndices.size(); j++) {
+				if (vocab[ii] == vocab[foundIndices[j]]) {
+					found = true;
+				}
+			}
+			if (!found) {
+				score_0 += std::log(1 - theta[0][vocab[ii]]);
+				score_1 += std::log(1 - theta[1][vocab[ii]]);
+			}
+		}
+
+		score_0 += std::log(prior_0);
+		score_1 += std::log(prior_1);
+
+		score_0 = exp(score_0);
+		score_1 = exp(score_1);
+
+		std::cout << score_0 << std::endl;
+		std::cout << score_1 << std::endl;
+
+		// Assigning the traning example to a class
+
+		if (score_0 > score_1) {
+			y_hat[i] = 0;
+		} else {
+			y_hat[i] = 1;
+		}
+	}
+}

mlpp/bernoulli_nb/bernoulli_nb_old.h

@@ -0,0 +1,42 @@
+
+#ifndef MLPP_BERNOULLI_NB_OLD_H
+#define MLPP_BERNOULLI_NB_OLD_H
+
+//
+//  BernoulliNB.hpp
+//
+//  Created by Marc Melikyan on 1/17/21.
+//
+
+#include "core/math/math_defs.h"
+
+#include <map>
+#include <vector>
+
+class MLPPBernoulliNBOld {
+public:
+	MLPPBernoulliNBOld(std::vector<std::vector<real_t>> inputSet, std::vector<real_t> outputSet);
+	std::vector<real_t> modelSetTest(std::vector<std::vector<real_t>> X);
+	real_t modelTest(std::vector<real_t> x);
+	real_t score();
+
+private:
+	void computeVocab();
+	void computeTheta();
+	void Evaluate();
+
+	// Model Params
+	real_t prior_1 = 0;
+	real_t prior_0 = 0;
+
+	std::vector<std::map<real_t, int>> theta;
+	std::vector<real_t> vocab;
+	int class_num;
+
+	// Datasets
+	std::vector<std::vector<real_t>> inputSet;
+	std::vector<real_t> outputSet;
+	std::vector<real_t> y_hat;
+};
+
+#endif /* BernoulliNB_hpp */

mlpp/c_log_log_reg/c_log_log_reg_old.cpp

@@ -0,0 +1,224 @@
+//
+//  CLogLogReg.cpp
+//
+//  Created by Marc Melikyan on 10/2/20.
+//
+
+#include "c_log_log_reg_old.h"
+
+#include "../activation/activation.h"
+#include "../cost/cost.h"
+#include "../lin_alg/lin_alg.h"
+#include "../regularization/reg.h"
+#include "../utilities/utilities.h"
+
+#include <iostream>
+#include <random>
+
+MLPPCLogLogRegOld::MLPPCLogLogRegOld(std::vector<std::vector<real_t>> inputSet, std::vector<real_t> outputSet, std::string reg, real_t lambda, real_t alpha) :
+		inputSet(inputSet), outputSet(outputSet), n(inputSet.size()), k(inputSet[0].size()), reg(reg), lambda(lambda), alpha(alpha) {
+	y_hat.resize(n);
+	weights = MLPPUtilities::weightInitialization(k);
+	bias = MLPPUtilities::biasInitialization();
+}
+
+std::vector<real_t> MLPPCLogLogRegOld::modelSetTest(std::vector<std::vector<real_t>> X) {
+	return Evaluate(X);
+}
+
+real_t MLPPCLogLogRegOld::modelTest(std::vector<real_t> x) {
+	return Evaluate(x);
+}
+
+void MLPPCLogLogRegOld::gradientDescent(real_t learning_rate, int max_epoch, bool UI) {
+	MLPPActivation avn;
+	MLPPLinAlg alg;
+	MLPPReg regularization;
+	real_t cost_prev = 0;
+	int epoch = 1;
+	forwardPass();
+
+	while (true) {
+		cost_prev = Cost(y_hat, outputSet);
+
+		std::vector<real_t> error = alg.subtraction(y_hat, outputSet);
+
+		// Calculating the weight gradients
+		weights = alg.subtraction(weights, alg.scalarMultiply(learning_rate / n, alg.mat_vec_mult(alg.transpose(inputSet), alg.hadamard_product(error, avn.cloglog(z, 1)))));
+		weights = regularization.regWeights(weights, lambda, alpha, reg);
+
+		// Calculating the bias gradients
+		bias -= learning_rate * alg.sum_elements(alg.hadamard_product(error, avn.cloglog(z, 1))) / n;
+
+		forwardPass();
+
+		if (UI) {
+			MLPPUtilities::CostInfo(epoch, cost_prev, Cost(y_hat, outputSet));
+			MLPPUtilities::UI(weights, bias);
+		}
+		epoch++;
+
+		if (epoch > max_epoch) {
+			break;
+		}
+	}
+}
+
+void MLPPCLogLogRegOld::MLE(real_t learning_rate, int max_epoch, bool UI) {
+	MLPPActivation avn;
+	MLPPLinAlg alg;
+	MLPPReg regularization;
+	real_t cost_prev = 0;
+	int epoch = 1;
+	forwardPass();
+
+	while (true) {
+		cost_prev = Cost(y_hat, outputSet);
+
+		std::vector<real_t> error = alg.subtraction(y_hat, outputSet);
+
+		weights = alg.addition(weights, alg.scalarMultiply(learning_rate / n, alg.mat_vec_mult(alg.transpose(inputSet), alg.hadamard_product(error, avn.cloglog(z, 1)))));
+		weights = regularization.regWeights(weights, lambda, alpha, reg);
+
+		// Calculating the bias gradients
+		bias += learning_rate * alg.sum_elements(alg.hadamard_product(error, avn.cloglog(z, 1))) / n;
+		forwardPass();
+
+		if (UI) {
+			MLPPUtilities::CostInfo(epoch, cost_prev, Cost(y_hat, outputSet));
+			MLPPUtilities::UI(weights, bias);
+		}
+		epoch++;
+
+		if (epoch > max_epoch) {
+			break;
+		}
+	}
+}
+
+void MLPPCLogLogRegOld::SGD(real_t learning_rate, int max_epoch, bool UI) {
+	MLPPLinAlg alg;
+	MLPPReg regularization;
+	real_t cost_prev = 0;
+	int epoch = 1;
+	forwardPass();
+
+	while (true) {
+		std::random_device rd;
+		std::default_random_engine generator(rd());
+		std::uniform_int_distribution<int> distribution(0, int(n - 1));
+		int outputIndex = distribution(generator);
+
+		real_t y_hat = Evaluate(inputSet[outputIndex]);
+		real_t z = propagate(inputSet[outputIndex]);
+		cost_prev = Cost({ y_hat }, { outputSet[outputIndex] });
+
+		real_t error = y_hat - outputSet[outputIndex];
+
+		// Weight Updation
+		weights = alg.subtraction(weights, alg.scalarMultiply(learning_rate * error * exp(z - exp(z)), inputSet[outputIndex]));
+		weights = regularization.regWeights(weights, lambda, alpha, reg);
+
+		// Bias updation
+		bias -= learning_rate * error * exp(z - exp(z));
+
+		y_hat = Evaluate({ inputSet[outputIndex] });
+
+		if (UI) {
+			MLPPUtilities::CostInfo(epoch, cost_prev, Cost({ y_hat }, { outputSet[outputIndex] }));
+			MLPPUtilities::UI(weights, bias);
+		}
+		epoch++;
+
+		if (epoch > max_epoch) {
+			break;
+		}
+	}
+	forwardPass();
+}
+
+void MLPPCLogLogRegOld::MBGD(real_t learning_rate, int max_epoch, int mini_batch_size, bool UI) {
+	MLPPActivation avn;
+	MLPPLinAlg alg;
+	MLPPReg regularization;
+	real_t cost_prev = 0;
+	int epoch = 1;
+
+	// Creating the mini-batches
+	int n_mini_batch = n / mini_batch_size;
+	auto batches = MLPPUtilities::createMiniBatches(inputSet, outputSet, n_mini_batch);
+	auto inputMiniBatches = std::get<0>(batches);
+	auto outputMiniBatches = std::get<1>(batches);
+
+	while (true) {
+		for (int i = 0; i < n_mini_batch; i++) {
+			std::vector<real_t> y_hat = Evaluate(inputMiniBatches[i]);
+			std::vector<real_t> z = propagate(inputMiniBatches[i]);
+			cost_prev = Cost(y_hat, outputMiniBatches[i]);
+
+			std::vector<real_t> error = alg.subtraction(y_hat, outputMiniBatches[i]);
+
+			// Calculating the weight gradients
+			weights = alg.subtraction(weights, alg.scalarMultiply(learning_rate / n, alg.mat_vec_mult(alg.transpose(inputMiniBatches[i]), alg.hadamard_product(error, avn.cloglog(z, 1)))));
+			weights = regularization.regWeights(weights, lambda, alpha, reg);
+
+			// Calculating the bias gradients
+			bias -= learning_rate * alg.sum_elements(alg.hadamard_product(error, avn.cloglog(z, 1))) / n;
+
+			forwardPass();
+
+			y_hat = Evaluate(inputMiniBatches[i]);
+
+			if (UI) {
+				MLPPUtilities::CostInfo(epoch, cost_prev, Cost(y_hat, outputMiniBatches[i]));
+				MLPPUtilities::UI(weights, bias);
+			}
+		}
+		epoch++;
+		if (epoch > max_epoch) {
+			break;
+		}
+	}
+	forwardPass();
+}
+
+real_t MLPPCLogLogRegOld::score() {
+	MLPPUtilities util;
+	return util.performance(y_hat, outputSet);
+}
+
+real_t MLPPCLogLogRegOld::Cost(std::vector<real_t> y_hat, std::vector<real_t> y) {
+	MLPPReg regularization;
+	class MLPPCost cost;
+	return cost.MSE(y_hat, y) + regularization.regTerm(weights, lambda, alpha, reg);
+}
+
+std::vector<real_t> MLPPCLogLogRegOld::Evaluate(std::vector<std::vector<real_t>> X) {
+	MLPPLinAlg alg;
+	MLPPActivation avn;
+	return avn.cloglog(alg.scalarAdd(bias, alg.mat_vec_mult(X, weights)));
+}
+
+std::vector<real_t> MLPPCLogLogRegOld::propagate(std::vector<std::vector<real_t>> X) {
+	MLPPLinAlg alg;
+	return alg.scalarAdd(bias, alg.mat_vec_mult(X, weights));
+}
+
+real_t MLPPCLogLogRegOld::Evaluate(std::vector<real_t> x) {
+	MLPPLinAlg alg;
+	MLPPActivation avn;
+	return avn.cloglog(alg.dot(weights, x) + bias);
+}
+
+real_t MLPPCLogLogRegOld::propagate(std::vector<real_t> x) {
+	MLPPLinAlg alg;
+	return alg.dot(weights, x) + bias;
+}
+
+// cloglog ( wTx + b )
+void MLPPCLogLogRegOld::forwardPass() {
+	MLPPActivation avn;
+
+	z = propagate(inputSet);
+	y_hat = avn.cloglog(z);
+}

mlpp/c_log_log_reg/c_log_log_reg_old.h

@@ -0,0 +1,54 @@
+
+#ifndef MLPP_C_LOG_LOG_REG_H
+#define MLPP_C_LOG_LOG_REG_H
+
+//
+//  CLogLogReg.hpp
+//
+//  Created by Marc Melikyan on 10/2/20.
+//
+
+#include "core/math/math_defs.h"
+
+#include <string>
+#include <vector>
+
+class MLPPCLogLogRegOld {
+public:
+	MLPPCLogLogRegOld(std::vector<std::vector<real_t>> inputSet, std::vector<real_t> outputSet, std::string reg = "None", real_t lambda = 0.5, real_t alpha = 0.5);
+	std::vector<real_t> modelSetTest(std::vector<std::vector<real_t>> X);
+	real_t modelTest(std::vector<real_t> x);
+	void gradientDescent(real_t learning_rate, int max_epoch, bool UI = false);
+	void MLE(real_t learning_rate, int max_epoch, bool UI = false);
+	void SGD(real_t learning_rate, int max_epoch, bool UI = false);
+	void MBGD(real_t learning_rate, int max_epoch, int mini_batch_size, bool UI = false);
+	real_t score();
+
+private:
+	void weightInitialization(int k);
+	void biasInitialization();
+	real_t Cost(std::vector<real_t> y_hat, std::vector<real_t> y);
+
+	std::vector<real_t> Evaluate(std::vector<std::vector<real_t>> X);
+	std::vector<real_t> propagate(std::vector<std::vector<real_t>> X);
+	real_t Evaluate(std::vector<real_t> x);
+	real_t propagate(std::vector<real_t> x);
+	void forwardPass();
+
+	std::vector<std::vector<real_t>> inputSet;
+	std::vector<real_t> outputSet;
+	std::vector<real_t> y_hat;
+	std::vector<real_t> z;
+	std::vector<real_t> weights;
+	real_t bias;
+
+	int n;
+	int k;
+
+	// Regularization Params
+	std::string reg;
+	real_t lambda;
+	real_t alpha; /* This is the controlling param for Elastic Net*/
+};
+
+#endif /* CLogLogReg_hpp */

mlpp/dual_svc/dual_svc_old.cpp

@@ -0,0 +1,246 @@
+//
+//  DualSVC.cpp
+//
+//  Created by Marc Melikyan on 10/2/20.
+//
+
+#include "dual_svc_old.h"
+#include "../activation/activation.h"
+#include "../cost/cost.h"
+#include "../lin_alg/lin_alg.h"
+#include "../regularization/reg.h"
+#include "../utilities/utilities.h"
+
+#include <iostream>
+#include <random>
+
+MLPPDualSVCOld::MLPPDualSVCOld(std::vector<std::vector<real_t>> p_inputSet, std::vector<real_t> p_outputSet, real_t p_C, std::string p_kernel) {
+	inputSet = p_inputSet;
+	outputSet = p_outputSet;
+	n = p_inputSet.size();
+	k = p_inputSet[0].size();
+	C = p_C;
+	kernel = p_kernel;
+
+	y_hat.resize(n);
+	bias = MLPPUtilities::biasInitialization();
+	alpha = MLPPUtilities::weightInitialization(n); // One alpha for all training examples, as per the lagrangian multipliers.
+	K = kernelFunction(inputSet, inputSet, kernel); // For now this is unused. When non-linear kernels are added, the K will be manipulated.
+}
+
+std::vector<real_t> MLPPDualSVCOld::modelSetTest(std::vector<std::vector<real_t>> X) {
+	return Evaluate(X);
+}
+
+real_t MLPPDualSVCOld::modelTest(std::vector<real_t> x) {
+	return Evaluate(x);
+}
+
+void MLPPDualSVCOld::gradientDescent(real_t learning_rate, int max_epoch, bool UI) {
+	class MLPPCost cost;
+	MLPPActivation avn;
+	MLPPLinAlg alg;
+	MLPPReg regularization;
+	real_t cost_prev = 0;
+	int epoch = 1;
+	forwardPass();
+
+	while (true) {
+		cost_prev = Cost(alpha, inputSet, outputSet);
+
+		alpha = alg.subtraction(alpha, alg.scalarMultiply(learning_rate, cost.dualFormSVMDeriv(alpha, inputSet, outputSet)));
+
+		alphaProjection();
+
+		// Calculating the bias
+		real_t biasGradient = 0;
+		for (uint32_t i = 0; i < alpha.size(); i++) {
+			real_t sum = 0;
+			if (alpha[i] < C && alpha[i] > 0) {
+				for (uint32_t j = 0; j < alpha.size(); j++) {
+					if (alpha[j] > 0) {
+						sum += alpha[j] * outputSet[j] * alg.dot(inputSet[j], inputSet[i]); // TO DO: DON'T forget to add non-linear kernelizations.
+					}
+				}
+			}
+			biasGradient = (1 - outputSet[i] * sum) / outputSet[i];
+			break;
+		}
+		bias -= biasGradient * learning_rate;
+
+		forwardPass();
+
+		// UI PORTION
+		if (UI) {
+			MLPPUtilities::CostInfo(epoch, cost_prev, Cost(alpha, inputSet, outputSet));
+			MLPPUtilities::UI(alpha, bias);
+			std::cout << score() << std::endl; // TO DO: DELETE THIS.
+		}
+		epoch++;
+
+		if (epoch > max_epoch) {
+			break;
+		}
+	}
+}
+
+// void MLPPDualSVCOld::SGD(real_t learning_rate, int max_epoch, bool UI){
+//     class MLPPCost cost;
+//     MLPPActivation avn;
+//     MLPPLinAlg alg;
+//     MLPPReg regularization;
+
+//     real_t cost_prev = 0;
+//     int epoch = 1;
+
+//     while(true){
+//         std::random_device rd;
+//         std::default_random_engine generator(rd());
+//         std::uniform_int_distribution<int> distribution(0, int(n - 1));
+//         int outputIndex = distribution(generator);
+
+//         cost_prev = Cost(alpha, inputSet[outputIndex], outputSet[outputIndex]);
+
+//         // Bias updation
+//         bias -= learning_rate * costDeriv;
+
+//         y_hat = Evaluate({inputSet[outputIndex]});
+
+//         if(UI) {
+//             MLPPUtilities::CostInfo(epoch, cost_prev, Cost(alpha));
+//             MLPPUtilities::UI(weights, bias);
+//         }
+//         epoch++;
+
+//         if(epoch > max_epoch) { break; }
+//     }
+//     forwardPass();
+// }
+
+// void MLPPDualSVCOld::MBGD(real_t learning_rate, int max_epoch, int mini_batch_size, bool UI){
+//     class MLPPCost cost;
+//     MLPPActivation avn;
+//     MLPPLinAlg alg;
+//     MLPPReg regularization;
+//     real_t cost_prev = 0;
+//     int epoch = 1;
+
+//     // Creating the mini-batches
+//     int n_mini_batch = n/mini_batch_size;
+//     auto [inputMiniBatches, outputMiniBatches] = MLPPUtilities::createMiniBatches(inputSet, outputSet, n_mini_batch);
+
+//     while(true){
+//         for(int i = 0; i < n_mini_batch; i++){
+//             std::vector<real_t> y_hat = Evaluate(inputMiniBatches[i]);
+//             std::vector<real_t> z = propagate(inputMiniBatches[i]);
+//             cost_prev = Cost(z, outputMiniBatches[i], weights, C);
+
+//             // Calculating the weight gradients
+//             weights = alg.subtraction(weights, alg.scalarMultiply(learning_rate/n, alg.mat_vec_mult(alg.transpose(inputMiniBatches[i]), cost.HingeLossDeriv(z, outputMiniBatches[i], C))));
+//             weights = regularization.regWeights(weights, learning_rate/n, 0, "Ridge");
+
+//             // Calculating the bias gradients
+//             bias -= learning_rate * alg.sum_elements(cost.HingeLossDeriv(y_hat, outputMiniBatches[i], C)) / n;
+
+//             forwardPass();
+
+//             y_hat = Evaluate(inputMiniBatches[i]);
+
+//             if(UI) {
+//                 MLPPUtilities::CostInfo(epoch, cost_prev, Cost(z, outputMiniBatches[i], weights, C));
+//                 MLPPUtilities::UI(weights, bias);
+//             }
+//         }
+//         epoch++;
+//         if(epoch > max_epoch) { break; }
+//     }
+//     forwardPass();
+// }
+
+real_t MLPPDualSVCOld::score() {
+	MLPPUtilities util;
+	return util.performance(y_hat, outputSet);
+}
+
+void MLPPDualSVCOld::save(std::string fileName) {
+	MLPPUtilities util;
+	util.saveParameters(fileName, alpha, bias);
+}
+
+real_t MLPPDualSVCOld::Cost(std::vector<real_t> alpha, std::vector<std::vector<real_t>> X, std::vector<real_t> y) {
+	class MLPPCost cost;
+	return cost.dualFormSVM(alpha, X, y);
+}
+
+std::vector<real_t> MLPPDualSVCOld::Evaluate(std::vector<std::vector<real_t>> X) {
+	MLPPActivation avn;
+	return avn.sign(propagate(X));
+}
+
+std::vector<real_t> MLPPDualSVCOld::propagate(std::vector<std::vector<real_t>> X) {
+	MLPPLinAlg alg;
+	std::vector<real_t> z;
+	for (uint32_t i = 0; i < X.size(); i++) {
+		real_t sum = 0;
+		for (uint32_t j = 0; j < alpha.size(); j++) {
+			if (alpha[j] != 0) {
+				sum += alpha[j] * outputSet[j] * alg.dot(inputSet[j], X[i]); // TO DO: DON'T forget to add non-linear kernelizations.
+			}
+		}
+		sum += bias;
+		z.push_back(sum);
+	}
+	return z;
+}
+
+real_t MLPPDualSVCOld::Evaluate(std::vector<real_t> x) {
+	MLPPActivation avn;
+	return avn.sign(propagate(x));
+}
+
+real_t MLPPDualSVCOld::propagate(std::vector<real_t> x) {
+	MLPPLinAlg alg;
+	real_t z = 0;
+	for (uint32_t j = 0; j < alpha.size(); j++) {
+		if (alpha[j] != 0) {
+			z += alpha[j] * outputSet[j] * alg.dot(inputSet[j], x); // TO DO: DON'T forget to add non-linear kernelizations.
+		}
+	}
+	z += bias;
+	return z;
+}
+
+void MLPPDualSVCOld::forwardPass() {
+	MLPPActivation avn;
+
+	z = propagate(inputSet);
+	y_hat = avn.sign(z);
+}
+
+void MLPPDualSVCOld::alphaProjection() {
+	for (uint32_t i = 0; i < alpha.size(); i++) {
+		if (alpha[i] > C) {
+			alpha[i] = C;
+		} else if (alpha[i] < 0) {
+			alpha[i] = 0;
+		}
+	}
+}
+
+real_t MLPPDualSVCOld::kernelFunction(std::vector<real_t> u, std::vector<real_t> v, std::string kernel) {
+	MLPPLinAlg alg;
+	if (kernel == "Linear") {
+		return alg.dot(u, v);
+	}
+
+	return 0;
+}
+
+std::vector<std::vector<real_t>> MLPPDualSVCOld::kernelFunction(std::vector<std::vector<real_t>> A, std::vector<std::vector<real_t>> B, std::string kernel) {
+	MLPPLinAlg alg;
+	if (kernel == "Linear") {
+		return alg.matmult(inputSet, alg.transpose(inputSet));
+	}
+
+	return std::vector<std::vector<real_t>>();
+}

mlpp/dual_svc/dual_svc_old.h

@@ -0,0 +1,69 @@
+
+#ifndef MLPP_DUAL_SVC_OLD_H
+#define MLPP_DUAL_SVC_OLD_H
+
+//
+//  DualSVC.hpp
+//
+//  Created by Marc Melikyan on 10/2/20.
+//
+// http://disp.ee.ntu.edu.tw/~pujols/Support%20Vector%20Machine.pdf
+// http://ciml.info/dl/v0_99/ciml-v0_99-ch11.pdf
+// Were excellent for the practical intution behind the dual formulation.
+
+#include "core/math/math_defs.h"
+
+#include <string>
+#include <vector>
+
+class MLPPDualSVCOld {
+public:
+	MLPPDualSVCOld(std::vector<std::vector<real_t>> inputSet, std::vector<real_t> outputSet, real_t C, std::string kernel = "Linear");
+	MLPPDualSVCOld(std::vector<std::vector<real_t>> inputSet, std::vector<real_t> outputSet, real_t C, std::string kernel, real_t p, real_t c);
+
+	std::vector<real_t> modelSetTest(std::vector<std::vector<real_t>> X);
+	real_t modelTest(std::vector<real_t> x);
+	void gradientDescent(real_t learning_rate, int max_epoch, bool UI = false);
+	void SGD(real_t learning_rate, int max_epoch, bool UI = false);
+	void MBGD(real_t learning_rate, int max_epoch, int mini_batch_size, bool UI = false);
+	real_t score();
+	void save(std::string fileName);
+
+private:
+	void init();
+
+	real_t Cost(std::vector<real_t> alpha, std::vector<std::vector<real_t>> X, std::vector<real_t> y);
+
+	std::vector<real_t> Evaluate(std::vector<std::vector<real_t>> X);
+	std::vector<real_t> propagate(std::vector<std::vector<real_t>> X);
+	real_t Evaluate(std::vector<real_t> x);
+	real_t propagate(std::vector<real_t> x);
+	void forwardPass();
+
+	void alphaProjection();
+
+	real_t kernelFunction(std::vector<real_t> v, std::vector<real_t> u, std::string kernel);
+	std::vector<std::vector<real_t>> kernelFunction(std::vector<std::vector<real_t>> U, std::vector<std::vector<real_t>> V, std::string kernel);
+
+	std::vector<std::vector<real_t>> inputSet;
+	std::vector<real_t> outputSet;
+	std::vector<real_t> z;
+	std::vector<real_t> y_hat;
+	real_t bias;
+
+	std::vector<real_t> alpha;
+	std::vector<std::vector<real_t>> K;
+
+	real_t C;
+	int n;
+	int k;
+
+	std::string kernel;
+	real_t p; // Poly
+	real_t c; // Poly
+
+	// UI Portion
+	void UI(int epoch, real_t cost_prev);
+};
+
+#endif /* DualSVC_hpp */

mlpp/exp_reg/exp_reg_old.cpp

@@ -0,0 +1,247 @@
+//
+//  ExpReg.cpp
+//
+//  Created by Marc Melikyan on 10/2/20.
+//
+
+#include "exp_reg_old.h"
+
+#include "../cost/cost.h"
+#include "../lin_alg/lin_alg.h"
+#include "../regularization/reg.h"
+#include "../stat/stat.h"
+#include "../utilities/utilities.h"
+
+#include <iostream>
+#include <random>
+
+MLPPExpRegOld::MLPPExpRegOld(std::vector<std::vector<real_t>> p_inputSet, std::vector<real_t> p_outputSet, std::string p_reg, real_t p_lambda, real_t p_alpha) {
+	inputSet = p_inputSet;
+	outputSet = p_outputSet;
+	n = p_inputSet.size();
+	k = p_inputSet[0].size();
+	reg = p_reg;
+	lambda = p_lambda;
+	alpha = p_alpha;
+
+	y_hat.resize(n);
+	weights = MLPPUtilities::weightInitialization(k);
+	initial = MLPPUtilities::weightInitialization(k);
+	bias = MLPPUtilities::biasInitialization();
+}
+
+std::vector<real_t> MLPPExpRegOld::modelSetTest(std::vector<std::vector<real_t>> X) {
+	return Evaluate(X);
+}
+
+real_t MLPPExpRegOld::modelTest(std::vector<real_t> x) {
+	return Evaluate(x);
+}
+
+void MLPPExpRegOld::gradientDescent(real_t learning_rate, int max_epoch, bool UI) {
+	MLPPLinAlg alg;
+	MLPPReg regularization;
+	real_t cost_prev = 0;
+	int epoch = 1;
+	forwardPass();
+
+	while (true) {
+		cost_prev = Cost(y_hat, outputSet);
+
+		std::vector<real_t> error = alg.subtraction(y_hat, outputSet);
+
+		for (int i = 0; i < k; i++) {
+			// Calculating the weight gradient
+			real_t sum = 0;
+			for (int j = 0; j < n; j++) {
+				sum += error[j] * inputSet[j][i] * std::pow(weights[i], inputSet[j][i] - 1);
+			}
+			real_t w_gradient = sum / n;
+
+			// Calculating the initial gradient
+			real_t sum2 = 0;
+			for (int j = 0; j < n; j++) {
+				sum2 += error[j] * std::pow(weights[i], inputSet[j][i]);
+			}
+
+			real_t i_gradient = sum2 / n;
+
+			// Weight/initial updation
+			weights[i] -= learning_rate * w_gradient;
+			initial[i] -= learning_rate * i_gradient;
+		}
+		weights = regularization.regWeights(weights, lambda, alpha, reg);
+
+		// Calculating the bias gradient
+		real_t sum = 0;
+		for (int j = 0; j < n; j++) {
+			sum += (y_hat[j] - outputSet[j]);
+		}
+		real_t b_gradient = sum / n;
+
+		// bias updation
+		bias -= learning_rate * b_gradient;
+		forwardPass();
+
+		if (UI) {
+			MLPPUtilities::CostInfo(epoch, cost_prev, Cost(y_hat, outputSet));
+			MLPPUtilities::UI(weights, bias);
+		}
+		epoch++;
+
+		if (epoch > max_epoch) {
+			break;
+		}
+	}
+}
+
+void MLPPExpRegOld::SGD(real_t learning_rate, int max_epoch, bool UI) {
+	MLPPReg regularization;
+	real_t cost_prev = 0;
+	int epoch = 1;
+
+	while (true) {
+		std::random_device rd;
+		std::default_random_engine generator(rd());
+		std::uniform_int_distribution<int> distribution(0, int(n - 1));
+		int outputIndex = distribution(generator);
+
+		real_t y_hat = Evaluate(inputSet[outputIndex]);
+		cost_prev = Cost({ y_hat }, { outputSet[outputIndex] });
+
+		for (int i = 0; i < k; i++) {
+			// Calculating the weight gradients
+
+			real_t w_gradient = (y_hat - outputSet[outputIndex]) * inputSet[outputIndex][i] * std::pow(weights[i], inputSet[outputIndex][i] - 1);
+			real_t i_gradient = (y_hat - outputSet[outputIndex]) * std::pow(weights[i], inputSet[outputIndex][i]);
+
+			// Weight/initial updation
+			weights[i] -= learning_rate * w_gradient;
+			initial[i] -= learning_rate * i_gradient;
+		}
+		weights = regularization.regWeights(weights, lambda, alpha, reg);
+
+		// Calculating the bias gradients
+		real_t b_gradient = (y_hat - outputSet[outputIndex]);
+
+		// Bias updation
+		bias -= learning_rate * b_gradient;
+		y_hat = Evaluate({ inputSet[outputIndex] });
+
+		if (UI) {
+			MLPPUtilities::CostInfo(epoch, cost_prev, Cost({ y_hat }, { outputSet[outputIndex] }));
+			MLPPUtilities::UI(weights, bias);
+		}
+		epoch++;
+
+		if (epoch > max_epoch) {
+			break;
+		}
+	}
+	forwardPass();
+}
+
+void MLPPExpRegOld::MBGD(real_t learning_rate, int max_epoch, int mini_batch_size, bool UI) {
+	MLPPLinAlg alg;
+	MLPPReg regularization;
+	real_t cost_prev = 0;
+	int epoch = 1;
+
+	// Creating the mini-batches
+	int n_mini_batch = n / mini_batch_size;
+	auto batches = MLPPUtilities::createMiniBatches(inputSet, outputSet, n_mini_batch);
+	auto inputMiniBatches = std::get<0>(batches);
+	auto outputMiniBatches = std::get<1>(batches);
+
+	while (true) {
+		for (int i = 0; i < n_mini_batch; i++) {
+			std::vector<real_t> y_hat = Evaluate(inputMiniBatches[i]);
+			cost_prev = Cost(y_hat, outputMiniBatches[i]);
+			std::vector<real_t> error = alg.subtraction(y_hat, outputMiniBatches[i]);
+
+			for (int j = 0; j < k; j++) {
+				// Calculating the weight gradient
+				real_t sum = 0;
+				for (uint32_t k = 0; k < outputMiniBatches[i].size(); k++) {
+					sum += error[k] * inputMiniBatches[i][k][j] * std::pow(weights[j], inputMiniBatches[i][k][j] - 1);
+				}
+				real_t w_gradient = sum / outputMiniBatches[i].size();
+
+				// Calculating the initial gradient
+				real_t sum2 = 0;
+				for (uint32_t k = 0; k < outputMiniBatches[i].size(); k++) {
+					sum2 += error[k] * std::pow(weights[j], inputMiniBatches[i][k][j]);
+				}
+
+				real_t i_gradient = sum2 / outputMiniBatches[i].size();
+
+				// Weight/initial updation
+				weights[j] -= learning_rate * w_gradient;
+				initial[j] -= learning_rate * i_gradient;
+			}
+			weights = regularization.regWeights(weights, lambda, alpha, reg);
+
+			// Calculating the bias gradient
+			real_t sum = 0;
+			for (uint32_t j = 0; j < outputMiniBatches[i].size(); j++) {
+				sum += (y_hat[j] - outputMiniBatches[i][j]);
+			}
+
+			//real_t b_gradient = sum / outputMiniBatches[i].size();
+			y_hat = Evaluate(inputMiniBatches[i]);
+
+			if (UI) {
+				MLPPUtilities::CostInfo(epoch, cost_prev, Cost(y_hat, outputMiniBatches[i]));
+				MLPPUtilities::UI(weights, bias);
+			}
+		}
+		epoch++;
+		if (epoch > max_epoch) {
+			break;
+		}
+	}
+	forwardPass();
+}
+
+real_t MLPPExpRegOld::score() {
+	MLPPUtilities util;
+	return util.performance(y_hat, outputSet);
+}
+
+void MLPPExpRegOld::save(std::string fileName) {
+	MLPPUtilities util;
+	util.saveParameters(fileName, weights, initial, bias);
+}
+
+real_t MLPPExpRegOld::Cost(std::vector<real_t> y_hat, std::vector<real_t> y) {
+	MLPPReg regularization;
+	class MLPPCost cost;
+	return cost.MSE(y_hat, y) + regularization.regTerm(weights, lambda, alpha, reg);
+}
+
+std::vector<real_t> MLPPExpRegOld::Evaluate(std::vector<std::vector<real_t>> X) {
+	std::vector<real_t> y_hat;
+	y_hat.resize(X.size());
+	for (uint32_t i = 0; i < X.size(); i++) {
+		y_hat[i] = 0;
+		for (uint32_t j = 0; j < X[i].size(); j++) {
+			y_hat[i] += initial[j] * std::pow(weights[j], X[i][j]);
+		}
+		y_hat[i] += bias;
+	}
+	return y_hat;
+}
+
+real_t MLPPExpRegOld::Evaluate(std::vector<real_t> x) {
+	real_t y_hat = 0;
+	for (uint32_t i = 0; i < x.size(); i++) {
+		y_hat += initial[i] * std::pow(weights[i], x[i]);
+	}
+
+	return y_hat + bias;
+}
+
+// a * w^x + b
+void MLPPExpRegOld::forwardPass() {
+	y_hat = Evaluate(inputSet);
+}

mlpp/exp_reg/exp_reg_old.h

@@ -0,0 +1,50 @@
+
+#ifndef MLPP_EXP_REG_OLD_H
+#define MLPP_EXP_REG_OLD_H
+
+//
+//  ExpReg.hpp
+//
+//  Created by Marc Melikyan on 10/2/20.
+//
+
+#include "core/math/math_defs.h"
+
+#include <string>
+#include <vector>
+
+class MLPPExpRegOld {
+public:
+	MLPPExpRegOld(std::vector<std::vector<real_t>> inputSet, std::vector<real_t> outputSet, std::string reg = "None", real_t lambda = 0.5, real_t alpha = 0.5);
+	std::vector<real_t> modelSetTest(std::vector<std::vector<real_t>> X);
+	real_t modelTest(std::vector<real_t> x);
+	void gradientDescent(real_t learning_rate, int max_epoch, bool UI = 1);
+	void SGD(real_t learning_rate, int max_epoch, bool UI = 1);
+	void MBGD(real_t learning_rate, int max_epoch, int mini_batch_size, bool UI = 1);
+	real_t score();
+	void save(std::string fileName);
+
+private:
+	real_t Cost(std::vector<real_t> y_hat, std::vector<real_t> y);
+
+	std::vector<real_t> Evaluate(std::vector<std::vector<real_t>> X);
+	real_t Evaluate(std::vector<real_t> x);
+	void forwardPass();
+
+	std::vector<std::vector<real_t>> inputSet;
+	std::vector<real_t> outputSet;
+	std::vector<real_t> y_hat;
+	std::vector<real_t> weights;
+	std::vector<real_t> initial;
+	real_t bias;
+
+	int n;
+	int k;
+
+	// Regularization Params
+	std::string reg;
+	real_t lambda;
+	real_t alpha; /* This is the controlling param for Elastic Net*/
+};
+
+#endif /* ExpReg_hpp */

mlpp/gan/gan.h

@@ -1,6 +1,6 @@
 
-#ifndef MLPP_GAN_hpp
-#define MLPP_GAN_hpp
+#ifndef MLPP_GAN_H
+#define MLPP_GAN_H
 
 //
 //  GAN.hpp

mlpp/gan/gan_old.cpp

@@ -0,0 +1,287 @@
+//
+//  GAN.cpp
+//
+//  Created by Marc Melikyan on 11/4/20.
+//
+
+#include "gan_old.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>
+
+MLPPGAN::MLPPGAN(real_t k, std::vector<std::vector<real_t>> outputSet) :
+		outputSet(outputSet), n(outputSet.size()), k(k) {
+}
+
+MLPPGAN::~MLPPGAN() {
+	delete outputLayer;
+}
+
+std::vector<std::vector<real_t>> MLPPGAN::generateExample(int n) {
+	MLPPLinAlg alg;
+	return modelSetTestGenerator(alg.gaussianNoise(n, k));
+}
+
+void MLPPGAN::gradientDescent(real_t learning_rate, int max_epoch, bool UI) {
+	class MLPPCost cost;
+	MLPPLinAlg alg;
+	real_t cost_prev = 0;
+	int epoch = 1;
+	forwardPass();
+
+	while (true) {
+		cost_prev = Cost(y_hat, alg.onevec(n));
+
+		// Training of the discriminator.
+
+		std::vector<std::vector<real_t>> generatorInputSet = alg.gaussianNoise(n, k);
+		std::vector<std::vector<real_t>> discriminatorInputSet = modelSetTestGenerator(generatorInputSet);
+		discriminatorInputSet.insert(discriminatorInputSet.end(), outputSet.begin(), outputSet.end()); // Fake + real inputs.
+
+		std::vector<real_t> y_hat = modelSetTestDiscriminator(discriminatorInputSet);
+		std::vector<real_t> outputSet = alg.zerovec(n);
+		std::vector<real_t> outputSetReal = alg.onevec(n);
+		outputSet.insert(outputSet.end(), outputSetReal.begin(), outputSetReal.end()); // Fake + real output scores.
+
+		auto dgrads = computeDiscriminatorGradients(y_hat, outputSet);
+		auto cumulativeDiscriminatorHiddenLayerWGrad = std::get<0>(dgrads);
+		auto outputDiscriminatorWGrad = std::get<1>(dgrads);
+
+		cumulativeDiscriminatorHiddenLayerWGrad = alg.scalarMultiply(learning_rate / n, cumulativeDiscriminatorHiddenLayerWGrad);
+		outputDiscriminatorWGrad = alg.scalarMultiply(learning_rate / n, outputDiscriminatorWGrad);
+		updateDiscriminatorParameters(cumulativeDiscriminatorHiddenLayerWGrad, outputDiscriminatorWGrad, learning_rate);
+
+		// Training of the generator.
+		generatorInputSet = alg.gaussianNoise(n, k);
+		discriminatorInputSet = modelSetTestGenerator(generatorInputSet);
+		y_hat = modelSetTestDiscriminator(discriminatorInputSet);
+		outputSet = alg.onevec(n);
+
+		std::vector<std::vector<std::vector<real_t>>> cumulativeGeneratorHiddenLayerWGrad = computeGeneratorGradients(y_hat, outputSet);
+		cumulativeGeneratorHiddenLayerWGrad = alg.scalarMultiply(learning_rate / n, cumulativeGeneratorHiddenLayerWGrad);
+		updateGeneratorParameters(cumulativeGeneratorHiddenLayerWGrad, learning_rate);
+
+		forwardPass();
+		if (UI) {
+			MLPPGAN::UI(epoch, cost_prev, MLPPGAN::y_hat, alg.onevec(n));
+		}
+
+		epoch++;
+		if (epoch > max_epoch) {
+			break;
+		}
+	}
+}
+
+real_t MLPPGAN::score() {
+	MLPPLinAlg alg;
+	MLPPUtilities util;
+	forwardPass();
+	return util.performance(y_hat, alg.onevec(n));
+}
+
+void MLPPGAN::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 MLPPGAN::addLayer(int n_hidden, std::string activation, std::string weightInit, std::string reg, real_t lambda, real_t alpha) {
+	MLPPLinAlg alg;
+	if (network.empty()) {
+		network.push_back(MLPPOldHiddenLayer(n_hidden, activation, alg.gaussianNoise(n, k), 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 MLPPGAN::addOutputLayer(std::string weightInit, std::string reg, real_t lambda, real_t alpha) {
+	MLPPLinAlg alg;
+	if (!network.empty()) {
+		outputLayer = new MLPPOldOutputLayer(network[network.size() - 1].n_hidden, "Sigmoid", "LogLoss", network[network.size() - 1].a, weightInit, reg, lambda, alpha);
+	} else {
+		outputLayer = new MLPPOldOutputLayer(k, "Sigmoid", "LogLoss", alg.gaussianNoise(n, k), weightInit, reg, lambda, alpha);
+	}
+}
+
+std::vector<std::vector<real_t>> MLPPGAN::modelSetTestGenerator(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() / 2; i++) {
+			network[i].input = network[i - 1].a;
+			network[i].forwardPass();
+		}
+	}
+	return network[network.size() / 2].a;
+}
+
+std::vector<real_t> MLPPGAN::modelSetTestDiscriminator(std::vector<std::vector<real_t>> X) {
+	if (!network.empty()) {
+		for (uint32_t i = network.size() / 2 + 1; i < network.size(); i++) {
+			if (i == network.size() / 2 + 1) {
+				network[i].input = X;
+			} else {
+				network[i].input = network[i - 1].a;
+			}
+			network[i].forwardPass();
+		}
+		outputLayer->input = network[network.size() - 1].a;
+	}
+	outputLayer->forwardPass();
+	return outputLayer->a;
+}
+
+real_t MLPPGAN::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 MLPPGAN::forwardPass() {
+	MLPPLinAlg alg;
+	if (!network.empty()) {
+		network[0].input = alg.gaussianNoise(n, k);
+		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 { // Should never happen, though.
+		outputLayer->input = alg.gaussianNoise(n, k);
+	}
+	outputLayer->forwardPass();
+	y_hat = outputLayer->a;
+}
+
+void MLPPGAN::updateDiscriminatorParameters(std::vector<std::vector<std::vector<real_t>>> hiddenLayerUpdations, std::vector<real_t> outputLayerUpdation, real_t learning_rate) {
+	MLPPLinAlg 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 = static_cast<int>(network.size()) - 2; i > static_cast<int>(network.size()) / 2; 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));
+		}
+	}
+}
+
+void MLPPGAN::updateGeneratorParameters(std::vector<std::vector<std::vector<real_t>>> hiddenLayerUpdations, real_t learning_rate) {
+	MLPPLinAlg alg;
+
+	if (!network.empty()) {
+		for (int i = network.size() / 2; i >= 0; i--) {
+			//std::cout << network[i].weights.size() << "x" << network[i].weights[0].size() << std::endl;
+			//std::cout << hiddenLayerUpdations[(network.size() - 2) - i + 1].size() << "x" << hiddenLayerUpdations[(network.size() - 2) - i + 1][0].size() << std::endl;
+			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>> MLPPGAN::computeDiscriminatorGradients(std::vector<real_t> y_hat, std::vector<real_t> outputSet) {
+	class MLPPCost cost;
+	MLPPActivation avn;
+	MLPPLinAlg 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.
+
+		//std::cout << "HIDDENLAYER FIRST:" << hiddenLayerWGrad.size() << "x" << hiddenLayerWGrad[0].size() << std::endl;
+		//std::cout << "WEIGHTS SECOND:" << network[network.size() - 1].weights.size() << "x" << network[network.size() - 1].weights[0].size() << std::endl;
+
+		for (int i = static_cast<int>(network.size()) - 2; i > static_cast<int>(network.size()) / 2; 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 };
+}
+
+std::vector<std::vector<std::vector<real_t>>> MLPPGAN::computeGeneratorGradients(std::vector<real_t> y_hat, std::vector<real_t> outputSet) {
+	class MLPPCost cost;
+	MLPPActivation avn;
+	MLPPLinAlg 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;
+}
+
+void MLPPGAN::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);
+		}
+	}
+}

mlpp/gan/gan_old.h

@@ -0,0 +1,59 @@
+
+#ifndef MLPP_GAN_OLD_hpp
+#define MLPP_GAN_OLD_hpp
+
+//
+//  GAN.hpp
+//
+//  Created by Marc Melikyan on 11/4/20.
+//
+
+#include "core/math/math_defs.h"
+
+#include "../hidden_layer/hidden_layer.h"
+#include "../output_layer/output_layer.h"
+
+#include "../hidden_layer/hidden_layer_old.h"
+#include "../output_layer/output_layer_old.h"
+
+#include <string>
+#include <tuple>
+#include <vector>
+
+class MLPPGAN {
+public:
+	MLPPGAN(real_t k, std::vector<std::vector<real_t>> outputSet);
+	~MLPPGAN();
+	std::vector<std::vector<real_t>> generateExample(int n);
+	void gradientDescent(real_t learning_rate, int max_epoch, bool UI = false);
+	real_t score();
+	void save(std::string fileName);
+
+	void addLayer(int n_hidden, std::string activation, std::string weightInit = "Default", std::string reg = "None", real_t lambda = 0.5, real_t alpha = 0.5);
+	void addOutputLayer(std::string weightInit = "Default", std::string reg = "None", real_t lambda = 0.5, real_t alpha = 0.5);
+
+private:
+	std::vector<std::vector<real_t>> modelSetTestGenerator(std::vector<std::vector<real_t>> X); // Evaluator for the generator of the gan.
+	std::vector<real_t> modelSetTestDiscriminator(std::vector<std::vector<real_t>> X); // Evaluator for the discriminator of the gan.
+
+	real_t Cost(std::vector<real_t> y_hat, std::vector<real_t> y);
+
+	void forwardPass();
+	void updateDiscriminatorParameters(std::vector<std::vector<std::vector<real_t>>> hiddenLayerUpdations, std::vector<real_t> outputLayerUpdation, real_t learning_rate);
+	void updateGeneratorParameters(std::vector<std::vector<std::vector<real_t>>> hiddenLayerUpdations, real_t learning_rate);
+	std::tuple<std::vector<std::vector<std::vector<real_t>>>, std::vector<real_t>> computeDiscriminatorGradients(std::vector<real_t> y_hat, std::vector<real_t> outputSet);
+	std::vector<std::vector<std::vector<real_t>>> computeGeneratorGradients(std::vector<real_t> y_hat, std::vector<real_t> outputSet);
+
+	void UI(int epoch, real_t cost_prev, std::vector<real_t> y_hat, std::vector<real_t> outputSet);
+
+	std::vector<std::vector<real_t>> outputSet;
+	std::vector<real_t> y_hat;
+
+	std::vector<MLPPOldHiddenLayer> network;
+	MLPPOldOutputLayer *outputLayer;
+
+	int n;
+	int k;
+};
+
+#endif /* GAN_hpp */

mlpp/gaussian_nb/gaussian_nb_old.cpp

@@ -0,0 +1,89 @@
+//
+//  GaussianNB.cpp
+//
+//  Created by Marc Melikyan on 1/17/21.
+//
+
+#include "gaussian_nb_old.h"
+
+#include "../lin_alg/lin_alg.h"
+#include "../stat/stat.h"
+#include "../utilities/utilities.h"
+
+#include <algorithm>
+#include <iostream>
+#include <random>
+
+MLPPGaussianNBOld::MLPPGaussianNBOld(std::vector<std::vector<real_t>> p_inputSet, std::vector<real_t> p_outputSet, int p_class_num) {
+	inputSet = p_inputSet;
+	outputSet = p_outputSet;
+	class_num = p_class_num;
+
+	y_hat.resize(outputSet.size());
+	Evaluate();
+}
+
+std::vector<real_t> MLPPGaussianNBOld::modelSetTest(std::vector<std::vector<real_t>> X) {
+	std::vector<real_t> y_hat;
+	for (uint32_t i = 0; i < X.size(); i++) {
+		y_hat.push_back(modelTest(X[i]));
+	}
+	return y_hat;
+}
+
+real_t MLPPGaussianNBOld::modelTest(std::vector<real_t> x) {
+	real_t score[class_num];
+	real_t y_hat_i = 1;
+	for (int i = class_num - 1; i >= 0; i--) {
+		y_hat_i += std::log(priors[i] * (1 / sqrt(2 * M_PI * sigma[i] * sigma[i])) * exp(-(x[i] * mu[i]) * (x[i] * mu[i]) / (2 * sigma[i] * sigma[i])));
+		score[i] = exp(y_hat_i);
+	}
+	return std::distance(score, std::max_element(score, score + sizeof(score) / sizeof(real_t)));
+}
+
+real_t MLPPGaussianNBOld::score() {
+	MLPPUtilities util;
+	return util.performance(y_hat, outputSet);
+}
+
+void MLPPGaussianNBOld::Evaluate() {
+	MLPPStat stat;
+	MLPPLinAlg alg;
+
+	// Computing mu_k_y and sigma_k_y
+	mu.resize(class_num);
+	sigma.resize(class_num);
+	for (int i = class_num - 1; i >= 0; i--) {
+		std::vector<real_t> set;
+		for (uint32_t j = 0; j < inputSet.size(); j++) {
+			for (uint32_t k = 0; k < inputSet[j].size(); k++) {
+				if (outputSet[j] == i) {
+					set.push_back(inputSet[j][k]);
+				}
+			}
+		}
+		mu[i] = stat.mean(set);
+		sigma[i] = stat.standardDeviation(set);
+	}
+
+	// Priors
+	priors.resize(class_num);
+	for (uint32_t i = 0; i < outputSet.size(); i++) {
+		priors[int(outputSet[i])]++;
+	}
+	priors = alg.scalarMultiply(real_t(1) / real_t(outputSet.size()), priors);
+
+	for (uint32_t i = 0; i < outputSet.size(); i++) {
+		real_t score[class_num];
+		real_t y_hat_i = 1;
+		for (int j = class_num - 1; j >= 0; j--) {
+			for (uint32_t k = 0; k < inputSet[i].size(); k++) {
+				y_hat_i += std::log(priors[j] * (1 / sqrt(2 * M_PI * sigma[j] * sigma[j])) * exp(-(inputSet[i][k] * mu[j]) * (inputSet[i][k] * mu[j]) / (2 * sigma[j] * sigma[j])));
+			}
+			score[j] = exp(y_hat_i);
+			std::cout << score[j] << std::endl;
+		}
+		y_hat[i] = std::distance(score, std::max_element(score, score + sizeof(score) / sizeof(real_t)));
+		std::cout << std::distance(score, std::max_element(score, score + sizeof(score) / sizeof(real_t))) << std::endl;
+	}
+}

mlpp/gaussian_nb/gaussian_nb_old.h

@@ -0,0 +1,37 @@
+
+#ifndef MLPP_GAUSSIAN_NB_OLD_H
+#define MLPP_GAUSSIAN_NB_OLD_H
+
+//
+//  GaussianNB.hpp
+//
+//  Created by Marc Melikyan on 1/17/21.
+//
+
+#include "core/math/math_defs.h"
+
+#include <vector>
+
+class MLPPGaussianNBOld {
+public:
+	MLPPGaussianNBOld(std::vector<std::vector<real_t>> inputSet, std::vector<real_t> outputSet, int class_num);
+	std::vector<real_t> modelSetTest(std::vector<std::vector<real_t>> X);
+	real_t modelTest(std::vector<real_t> x);
+	real_t score();
+
+private:
+	void Evaluate();
+
+	int class_num;
+
+	std::vector<real_t> priors;
+	std::vector<real_t> mu;
+	std::vector<real_t> sigma;
+
+	std::vector<std::vector<real_t>> inputSet;
+	std::vector<real_t> outputSet;
+
+	std::vector<real_t> y_hat;
+};
+
+#endif /* GaussianNB_hpp */

mlpp/lin_reg/lin_reg_old.cpp

@@ -0,0 +1,598 @@
+//
+//  LinReg.cpp
+//
+//  Created by Marc Melikyan on 10/2/20.
+//
+
+#include "lin_reg_old.h"
+
+#include "../cost/cost.h"
+#include "../lin_alg/lin_alg.h"
+#include "../regularization/reg.h"
+#include "../stat/stat.h"
+#include "../utilities/utilities.h"
+
+#include <cmath>
+#include <iostream>
+#include <random>
+
+MLPPLinRegOld::MLPPLinRegOld(std::vector<std::vector<real_t>> p_inputSet, std::vector<real_t> p_outputSet, std::string p_reg, real_t p_lambda, real_t p_alpha) {
+	inputSet = p_inputSet;
+	outputSet = p_outputSet;
+	n = p_inputSet.size();
+	k = p_inputSet[0].size();
+	reg = p_reg;
+	lambda = p_lambda;
+	alpha = p_alpha;
+
+	y_hat.resize(n);
+
+	weights = MLPPUtilities::weightInitialization(k);
+	bias = MLPPUtilities::biasInitialization();
+}
+
+std::vector<real_t> MLPPLinRegOld::modelSetTest(std::vector<std::vector<real_t>> X) {
+	return Evaluate(X);
+}
+
+real_t MLPPLinRegOld::modelTest(std::vector<real_t> x) {
+	return Evaluate(x);
+}
+
+void MLPPLinRegOld::NewtonRaphson(real_t learning_rate, int max_epoch, bool UI) {
+	MLPPLinAlg alg;
+	MLPPReg regularization;
+	real_t cost_prev = 0;
+	int epoch = 1;
+	forwardPass();
+	while (true) {
+		cost_prev = Cost(y_hat, outputSet);
+
+		std::vector<real_t> error = alg.subtraction(y_hat, outputSet);
+
+		// Calculating the weight gradients (2nd derivative)
+		std::vector<real_t> first_derivative = alg.mat_vec_mult(alg.transpose(inputSet), error);
+		std::vector<std::vector<real_t>> second_derivative = alg.matmult(alg.transpose(inputSet), inputSet);
+		weights = alg.subtraction(weights, alg.scalarMultiply(learning_rate / n, alg.mat_vec_mult(alg.transpose(alg.inverse(second_derivative)), first_derivative)));
+		weights = regularization.regWeights(weights, lambda, alpha, reg);
+
+		// Calculating the bias gradients (2nd derivative)
+		bias -= learning_rate * alg.sum_elements(error) / n; // We keep this the same. The 2nd derivative is just [1].
+		forwardPass();
+
+		if (UI) {
+			MLPPUtilities::CostInfo(epoch, cost_prev, Cost(y_hat, outputSet));
+			MLPPUtilities::UI(weights, bias);
+		}
+		epoch++;
+		if (epoch > max_epoch) {
+			break;
+		}
+	}
+}
+
+void MLPPLinRegOld::gradientDescent(real_t learning_rate, int max_epoch, bool UI) {
+	MLPPLinAlg alg;
+	MLPPReg regularization;
+	real_t cost_prev = 0;
+	int epoch = 1;
+	forwardPass();
+
+	while (true) {
+		cost_prev = Cost(y_hat, outputSet);
+
+		std::vector<real_t> error = alg.subtraction(y_hat, outputSet);
+
+		// Calculating the weight gradients
+		weights = alg.subtraction(weights, alg.scalarMultiply(learning_rate / n, alg.mat_vec_mult(alg.transpose(inputSet), error)));
+		weights = regularization.regWeights(weights, lambda, alpha, reg);
+
+		// Calculating the bias gradients
+		bias -= learning_rate * alg.sum_elements(error) / n;
+		forwardPass();
+
+		if (UI) {
+			MLPPUtilities::CostInfo(epoch, cost_prev, Cost(y_hat, outputSet));
+			MLPPUtilities::UI(weights, bias);
+		}
+		epoch++;
+		if (epoch > max_epoch) {
+			break;
+		}
+	}
+}
+
+void MLPPLinRegOld::SGD(real_t learning_rate, int max_epoch, bool UI) {
+	MLPPLinAlg alg;
+	MLPPReg regularization;
+	real_t cost_prev = 0;
+	int epoch = 1;
+
+	while (true) {
+		std::random_device rd;
+		std::default_random_engine generator(rd());
+		std::uniform_int_distribution<int> distribution(0, int(n - 1));
+		int outputIndex = distribution(generator);
+
+		real_t y_hat = Evaluate(inputSet[outputIndex]);
+		cost_prev = Cost({ y_hat }, { outputSet[outputIndex] });
+
+		real_t error = y_hat - outputSet[outputIndex];
+
+		// Weight updation
+		weights = alg.subtraction(weights, alg.scalarMultiply(learning_rate * error, inputSet[outputIndex]));
+		weights = regularization.regWeights(weights, lambda, alpha, reg);
+
+		// Bias updation
+		bias -= learning_rate * error;
+
+		y_hat = Evaluate({ inputSet[outputIndex] });
+
+		if (UI) {
+			MLPPUtilities::CostInfo(epoch, cost_prev, Cost({ y_hat }, { outputSet[outputIndex] }));
+			MLPPUtilities::UI(weights, bias);
+		}
+		epoch++;
+
+		if (epoch > max_epoch) {
+			break;
+		}
+	}
+	forwardPass();
+}
+
+void MLPPLinRegOld::MBGD(real_t learning_rate, int max_epoch, int mini_batch_size, bool UI) {
+	MLPPLinAlg alg;
+	MLPPReg regularization;
+	real_t cost_prev = 0;
+	int epoch = 1;
+
+	// Creating the mini-batches
+	int n_mini_batch = n / mini_batch_size;
+	auto batches = MLPPUtilities::createMiniBatches(inputSet, outputSet, n_mini_batch);
+	auto inputMiniBatches = std::get<0>(batches);
+	auto outputMiniBatches = std::get<1>(batches);
+
+	while (true) {
+		for (int i = 0; i < n_mini_batch; i++) {
+			std::vector<real_t> y_hat = Evaluate(inputMiniBatches[i]);
+			cost_prev = Cost(y_hat, outputMiniBatches[i]);
+
+			std::vector<real_t> error = alg.subtraction(y_hat, outputMiniBatches[i]);
+
+			// Calculating the weight gradients
+			weights = alg.subtraction(weights, alg.scalarMultiply(learning_rate / outputMiniBatches[i].size(), alg.mat_vec_mult(alg.transpose(inputMiniBatches[i]), error)));
+			weights = regularization.regWeights(weights, lambda, alpha, reg);
+
+			// Calculating the bias gradients
+			bias -= learning_rate * alg.sum_elements(error) / outputMiniBatches[i].size();
+			y_hat = Evaluate(inputMiniBatches[i]);
+
+			if (UI) {
+				MLPPUtilities::CostInfo(epoch, cost_prev, Cost(y_hat, outputMiniBatches[i]));
+				MLPPUtilities::UI(weights, bias);
+			}
+		}
+		epoch++;
+		if (epoch > max_epoch) {
+			break;
+		}
+	}
+	forwardPass();
+}
+
+void MLPPLinRegOld::Momentum(real_t learning_rate, int max_epoch, int mini_batch_size, real_t gamma, bool UI) {
+	MLPPLinAlg alg;
+	MLPPReg regularization;
+	real_t cost_prev = 0;
+	int epoch = 1;
+
+	// Creating the mini-batches
+	int n_mini_batch = n / mini_batch_size;
+	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 Momentum.
+	std::vector<real_t> v = alg.zerovec(weights.size());
+	while (true) {
+		for (int i = 0; i < n_mini_batch; i++) {
+			std::vector<real_t> y_hat = Evaluate(inputMiniBatches[i]);
+			cost_prev = Cost(y_hat, outputMiniBatches[i]);
+
+			std::vector<real_t> error = alg.subtraction(y_hat, outputMiniBatches[i]);
+
+			// Calculating the weight gradients
+			std::vector<real_t> gradient = alg.scalarMultiply(1 / outputMiniBatches[i].size(), alg.mat_vec_mult(alg.transpose(inputMiniBatches[i]), error));
+			std::vector<real_t> RegDerivTerm = regularization.regDerivTerm(weights, lambda, alpha, reg);
+			std::vector<real_t> weight_grad = alg.addition(gradient, RegDerivTerm); // Weight_grad_final
+
+			v = alg.addition(alg.scalarMultiply(gamma, v), alg.scalarMultiply(learning_rate, weight_grad));
+
+			weights = alg.subtraction(weights, v);
+
+			// Calculating the bias gradients
+			bias -= learning_rate * alg.sum_elements(error) / outputMiniBatches[i].size(); // As normal
+			y_hat = Evaluate(inputMiniBatches[i]);
+
+			if (UI) {
+				MLPPUtilities::CostInfo(epoch, cost_prev, Cost(y_hat, outputMiniBatches[i]));
+				MLPPUtilities::UI(weights, bias);
+			}
+		}
+		epoch++;
+		if (epoch > max_epoch) {
+			break;
+		}
+	}
+	forwardPass();
+}
+
+void MLPPLinRegOld::NAG(real_t learning_rate, int max_epoch, int mini_batch_size, real_t gamma, bool UI) {
+	MLPPLinAlg alg;
+	MLPPReg regularization;
+	real_t cost_prev = 0;
+	int epoch = 1;
+
+	// Creating the mini-batches
+	int n_mini_batch = n / mini_batch_size;
+	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 Momentum.
+	std::vector<real_t> v = alg.zerovec(weights.size());
+	while (true) {
+		for (int i = 0; i < n_mini_batch; i++) {
+			weights = alg.subtraction(weights, alg.scalarMultiply(gamma, v)); // "Aposterori" calculation
+
+			std::vector<real_t> y_hat = Evaluate(inputMiniBatches[i]);
+			cost_prev = Cost(y_hat, outputMiniBatches[i]);
+
+			std::vector<real_t> error = alg.subtraction(y_hat, outputMiniBatches[i]);
+
+			// Calculating the weight gradients
+			std::vector<real_t> gradient = alg.scalarMultiply(1 / outputMiniBatches[i].size(), alg.mat_vec_mult(alg.transpose(inputMiniBatches[i]), error));
+			std::vector<real_t> RegDerivTerm = regularization.regDerivTerm(weights, lambda, alpha, reg);
+			std::vector<real_t> weight_grad = alg.addition(gradient, RegDerivTerm); // Weight_grad_final
+
+			v = alg.addition(alg.scalarMultiply(gamma, v), alg.scalarMultiply(learning_rate, weight_grad));
+
+			weights = alg.subtraction(weights, v);
+
+			// Calculating the bias gradients
+			bias -= learning_rate * alg.sum_elements(error) / outputMiniBatches[i].size(); // As normal
+			y_hat = Evaluate(inputMiniBatches[i]);
+
+			if (UI) {
+				MLPPUtilities::CostInfo(epoch, cost_prev, Cost(y_hat, outputMiniBatches[i]));
+				MLPPUtilities::UI(weights, bias);
+			}
+		}
+		epoch++;
+		if (epoch > max_epoch) {
+			break;
+		}
+	}
+	forwardPass();
+}
+
+void MLPPLinRegOld::Adagrad(real_t learning_rate, int max_epoch, int mini_batch_size, real_t e, bool UI) {
+	MLPPLinAlg alg;
+	MLPPReg regularization;
+	real_t cost_prev = 0;
+	int epoch = 1;
+
+	// Creating the mini-batches
+	int n_mini_batch = n / mini_batch_size;
+	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 Adagrad.
+	std::vector<real_t> v = alg.zerovec(weights.size());
+	while (true) {
+		for (int i = 0; i < n_mini_batch; i++) {
+			std::vector<real_t> y_hat = Evaluate(inputMiniBatches[i]);
+			cost_prev = Cost(y_hat, outputMiniBatches[i]);
+
+			std::vector<real_t> error = alg.subtraction(y_hat, outputMiniBatches[i]);
+
+			// Calculating the weight gradients
+			std::vector<real_t> gradient = alg.scalarMultiply(1 / outputMiniBatches[i].size(), alg.mat_vec_mult(alg.transpose(inputMiniBatches[i]), error));
+			std::vector<real_t> RegDerivTerm = regularization.regDerivTerm(weights, lambda, alpha, reg);
+			std::vector<real_t> weight_grad = alg.addition(gradient, RegDerivTerm); // Weight_grad_final
+
+			v = alg.hadamard_product(weight_grad, weight_grad);
+
+			weights = alg.subtraction(weights, alg.scalarMultiply(learning_rate, alg.elementWiseDivision(weight_grad, alg.sqrt(alg.scalarAdd(e, v)))));
+
+			// Calculating the bias gradients
+			bias -= learning_rate * alg.sum_elements(error) / outputMiniBatches[i].size(); // As normal
+			y_hat = Evaluate(inputMiniBatches[i]);
+
+			if (UI) {
+				MLPPUtilities::CostInfo(epoch, cost_prev, Cost(y_hat, outputMiniBatches[i]));
+				MLPPUtilities::UI(weights, bias);
+			}
+		}
+		epoch++;
+		if (epoch > max_epoch) {
+			break;
+		}
+	}
+	forwardPass();
+}
+
+void MLPPLinRegOld::Adadelta(real_t learning_rate, int max_epoch, int mini_batch_size, real_t b1, real_t e, bool UI) {
+	// Adagrad upgrade. Momentum is applied.
+	MLPPLinAlg alg;
+	MLPPReg regularization;
+	real_t cost_prev = 0;
+	int epoch = 1;
+
+	// Creating the mini-batches
+	int n_mini_batch = n / mini_batch_size;
+	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 Adagrad.
+	std::vector<real_t> v = alg.zerovec(weights.size());
+	while (true) {
+		for (int i = 0; i < n_mini_batch; i++) {
+			std::vector<real_t> y_hat = Evaluate(inputMiniBatches[i]);
+			cost_prev = Cost(y_hat, outputMiniBatches[i]);
+
+			std::vector<real_t> error = alg.subtraction(y_hat, outputMiniBatches[i]);
+
+			// Calculating the weight gradients
+			std::vector<real_t> gradient = alg.scalarMultiply(1 / outputMiniBatches[i].size(), alg.mat_vec_mult(alg.transpose(inputMiniBatches[i]), error));
+			std::vector<real_t> RegDerivTerm = regularization.regDerivTerm(weights, lambda, alpha, reg);
+			std::vector<real_t> weight_grad = alg.addition(gradient, RegDerivTerm); // Weight_grad_final
+
+			v = alg.addition(alg.scalarMultiply(b1, v), alg.scalarMultiply(1 - b1, alg.hadamard_product(weight_grad, weight_grad)));
+
+			weights = alg.subtraction(weights, alg.scalarMultiply(learning_rate, alg.elementWiseDivision(weight_grad, alg.sqrt(alg.scalarAdd(e, v)))));
+
+			// Calculating the bias gradients
+			bias -= learning_rate * alg.sum_elements(error) / outputMiniBatches[i].size(); // As normal
+			y_hat = Evaluate(inputMiniBatches[i]);
+
+			if (UI) {
+				MLPPUtilities::CostInfo(epoch, cost_prev, Cost(y_hat, outputMiniBatches[i]));
+				MLPPUtilities::UI(weights, bias);
+			}
+		}
+		epoch++;
+		if (epoch > max_epoch) {
+			break;
+		}
+	}
+	forwardPass();
+}
+
+void MLPPLinRegOld::Adam(real_t learning_rate, int max_epoch, int mini_batch_size, real_t b1, real_t b2, real_t e, bool UI) {
+	MLPPLinAlg alg;
+	MLPPReg regularization;
+	real_t cost_prev = 0;
+	int epoch = 1;
+
+	// Creating the mini-batches
+	int n_mini_batch = n / mini_batch_size;
+	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<real_t> m = alg.zerovec(weights.size());
+
+	std::vector<real_t> v = alg.zerovec(weights.size());
+	while (true) {
+		for (int i = 0; i < n_mini_batch; i++) {
+			std::vector<real_t> y_hat = Evaluate(inputMiniBatches[i]);
+			cost_prev = Cost(y_hat, outputMiniBatches[i]);
+
+			std::vector<real_t> error = alg.subtraction(y_hat, outputMiniBatches[i]);
+
+			// Calculating the weight gradients
+			std::vector<real_t> gradient = alg.scalarMultiply(1 / outputMiniBatches[i].size(), alg.mat_vec_mult(alg.transpose(inputMiniBatches[i]), error));
+			std::vector<real_t> RegDerivTerm = regularization.regDerivTerm(weights, lambda, alpha, reg);
+			std::vector<real_t> weight_grad = alg.addition(gradient, RegDerivTerm); // Weight_grad_final
+
+			m = alg.addition(alg.scalarMultiply(b1, m), alg.scalarMultiply(1 - b1, weight_grad));
+			v = alg.addition(alg.scalarMultiply(b2, v), alg.scalarMultiply(1 - b2, alg.exponentiate(weight_grad, 2)));
+
+			std::vector<real_t> m_hat = alg.scalarMultiply(1 / (1 - pow(b1, epoch)), m);
+			std::vector<real_t> v_hat = alg.scalarMultiply(1 / (1 - pow(b2, epoch)), v);
+
+			weights = alg.subtraction(weights, alg.scalarMultiply(learning_rate, alg.elementWiseDivision(m_hat, alg.scalarAdd(e, alg.sqrt(v_hat)))));
+
+			// Calculating the bias gradients
+			bias -= learning_rate * alg.sum_elements(error) / outputMiniBatches[i].size(); // As normal
+			y_hat = Evaluate(inputMiniBatches[i]);
+
+			if (UI) {
+				MLPPUtilities::CostInfo(epoch, cost_prev, Cost(y_hat, outputMiniBatches[i]));
+				MLPPUtilities::UI(weights, bias);
+			}
+		}
+		epoch++;
+		if (epoch > max_epoch) {
+			break;
+		}
+	}
+	forwardPass();
+}
+
+void MLPPLinRegOld::Adamax(real_t learning_rate, int max_epoch, int mini_batch_size, real_t b1, real_t b2, real_t e, bool UI) {
+	MLPPLinAlg alg;
+	MLPPReg regularization;
+	real_t cost_prev = 0;
+	int epoch = 1;
+
+	// Creating the mini-batches
+	int n_mini_batch = n / mini_batch_size;
+	auto batches = MLPPUtilities::createMiniBatches(inputSet, outputSet, n_mini_batch);
+	auto inputMiniBatches = std::get<0>(batches);
+	auto outputMiniBatches = std::get<1>(batches);
+
+	std::vector<real_t> m = alg.zerovec(weights.size());
+
+	std::vector<real_t> u = alg.zerovec(weights.size());
+	while (true) {
+		for (int i = 0; i < n_mini_batch; i++) {
+			std::vector<real_t> y_hat = Evaluate(inputMiniBatches[i]);
+			cost_prev = Cost(y_hat, outputMiniBatches[i]);
+
+			std::vector<real_t> error = alg.subtraction(y_hat, outputMiniBatches[i]);
+
+			// Calculating the weight gradients
+			std::vector<real_t> gradient = alg.scalarMultiply(1 / outputMiniBatches[i].size(), alg.mat_vec_mult(alg.transpose(inputMiniBatches[i]), error));
+			std::vector<real_t> RegDerivTerm = regularization.regDerivTerm(weights, lambda, alpha, reg);
+			std::vector<real_t> weight_grad = alg.addition(gradient, RegDerivTerm); // Weight_grad_final
+
+			m = alg.addition(alg.scalarMultiply(b1, m), alg.scalarMultiply(1 - b1, weight_grad));
+			u = alg.max(alg.scalarMultiply(b2, u), alg.abs(weight_grad));
+
+			std::vector<real_t> m_hat = alg.scalarMultiply(1 / (1 - pow(b1, epoch)), m);
+
+			weights = alg.subtraction(weights, alg.scalarMultiply(learning_rate, alg.elementWiseDivision(m_hat, u)));
+
+			// Calculating the bias gradients
+			bias -= learning_rate * alg.sum_elements(error) / outputMiniBatches[i].size(); // As normal
+			y_hat = Evaluate(inputMiniBatches[i]);
+
+			if (UI) {
+				MLPPUtilities::CostInfo(epoch, cost_prev, Cost(y_hat, outputMiniBatches[i]));
+				MLPPUtilities::UI(weights, bias);
+			}
+		}
+		epoch++;
+		if (epoch > max_epoch) {
+			break;
+		}
+	}
+	forwardPass();
+}
+
+void MLPPLinRegOld::Nadam(real_t learning_rate, int max_epoch, int mini_batch_size, real_t b1, real_t b2, real_t e, bool UI) {
+	MLPPLinAlg alg;
+	MLPPReg regularization;
+	real_t cost_prev = 0;
+	int epoch = 1;
+
+	// Creating the mini-batches
+	int n_mini_batch = n / mini_batch_size;
+	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<real_t> m = alg.zerovec(weights.size());
+	std::vector<real_t> v = alg.zerovec(weights.size());
+	std::vector<real_t> m_final = alg.zerovec(weights.size());
+	while (true) {
+		for (int i = 0; i < n_mini_batch; i++) {
+			std::vector<real_t> y_hat = Evaluate(inputMiniBatches[i]);
+			cost_prev = Cost(y_hat, outputMiniBatches[i]);
+
+			std::vector<real_t> error = alg.subtraction(y_hat, outputMiniBatches[i]);
+
+			// Calculating the weight gradients
+			std::vector<real_t> gradient = alg.scalarMultiply(1 / outputMiniBatches[i].size(), alg.mat_vec_mult(alg.transpose(inputMiniBatches[i]), error));
+			std::vector<real_t> RegDerivTerm = regularization.regDerivTerm(weights, lambda, alpha, reg);
+			std::vector<real_t> weight_grad = alg.addition(gradient, RegDerivTerm); // Weight_grad_final
+
+			m = alg.addition(alg.scalarMultiply(b1, m), alg.scalarMultiply(1 - b1, weight_grad));
+			v = alg.addition(alg.scalarMultiply(b2, v), alg.scalarMultiply(1 - b2, alg.exponentiate(weight_grad, 2)));
+			m_final = alg.addition(alg.scalarMultiply(b1, m), alg.scalarMultiply((1 - b1) / (1 - pow(b1, epoch)), weight_grad));
+
+			std::vector<real_t> m_hat = alg.scalarMultiply(1 / (1 - pow(b1, epoch)), m);
+			std::vector<real_t> v_hat = alg.scalarMultiply(1 / (1 - pow(b2, epoch)), v);
+
+			weights = alg.subtraction(weights, alg.scalarMultiply(learning_rate, alg.elementWiseDivision(m_final, alg.scalarAdd(e, alg.sqrt(v_hat)))));
+
+			// Calculating the bias gradients
+			bias -= learning_rate * alg.sum_elements(error) / outputMiniBatches[i].size(); // As normal
+			y_hat = Evaluate(inputMiniBatches[i]);
+
+			if (UI) {
+				MLPPUtilities::CostInfo(epoch, cost_prev, Cost(y_hat, outputMiniBatches[i]));
+				MLPPUtilities::UI(weights, bias);
+			}
+		}
+		epoch++;
+		if (epoch > max_epoch) {
+			break;
+		}
+	}
+	forwardPass();
+}
+
+void MLPPLinRegOld::normalEquation() {
+	MLPPLinAlg alg;
+	MLPPStat stat;
+	std::vector<real_t> x_means;
+	std::vector<std::vector<real_t>> inputSetT = alg.transpose(inputSet);
+
+	x_means.resize(inputSetT.size());
+	for (uint32_t i = 0; i < inputSetT.size(); i++) {
+		x_means[i] = (stat.mean(inputSetT[i]));
+	}
+
+	//try {
+	std::vector<real_t> temp;
+	temp.resize(k);
+	temp = alg.mat_vec_mult(alg.inverse(alg.matmult(alg.transpose(inputSet), inputSet)), alg.mat_vec_mult(alg.transpose(inputSet), outputSet));
+	if (std::isnan(temp[0])) {
+		//throw 99;
+		//TODO ERR_FAIL_COND
+		std::cout << "ERR: Resulting matrix was noninvertible/degenerate, and so the normal equation could not be performed. Try utilizing gradient descent." << std::endl;
+		return;
+	} else {
+		if (reg == "Ridge") {
+			weights = alg.mat_vec_mult(alg.inverse(alg.addition(alg.matmult(alg.transpose(inputSet), inputSet), alg.scalarMultiply(lambda, alg.identity(k)))), alg.mat_vec_mult(alg.transpose(inputSet), outputSet));
+		} else {
+			weights = alg.mat_vec_mult(alg.inverse(alg.matmult(alg.transpose(inputSet), inputSet)), alg.mat_vec_mult(alg.transpose(inputSet), outputSet));
+		}
+
+		bias = stat.mean(outputSet) - alg.dot(weights, x_means);
+
+		forwardPass();
+	}
+	//} catch (int err_num) {
+	//	std::cout << "ERR " << err_num << ": Resulting matrix was noninvertible/degenerate, and so the normal equation could not be performed. Try utilizing gradient descent." << std::endl;
+	//}
+}
+
+real_t MLPPLinRegOld::score() {
+	MLPPUtilities util;
+	return util.performance(y_hat, outputSet);
+}
+
+void MLPPLinRegOld::save(std::string fileName) {
+	MLPPUtilities util;
+	util.saveParameters(fileName, weights, bias);
+}
+
+real_t MLPPLinRegOld::Cost(std::vector<real_t> y_hat, std::vector<real_t> y) {
+	MLPPReg regularization;
+	class MLPPCost cost;
+	return cost.MSE(y_hat, y) + regularization.regTerm(weights, lambda, alpha, reg);
+}
+
+std::vector<real_t> MLPPLinRegOld::Evaluate(std::vector<std::vector<real_t>> X) {
+	MLPPLinAlg alg;
+	return alg.scalarAdd(bias, alg.mat_vec_mult(X, weights));
+}
+
+real_t MLPPLinRegOld::Evaluate(std::vector<real_t> x) {
+	MLPPLinAlg alg;
+	return alg.dot(weights, x) + bias;
+}
+
+// wTx + b
+void MLPPLinRegOld::forwardPass() {
+	y_hat = Evaluate(inputSet);
+}

mlpp/lin_reg/lin_reg_old.h

@@ -0,0 +1,60 @@
+
+#ifndef MLPP_LIN_REG_OLD_H
+#define MLPP_LIN_REG_OLD_H
+
+//
+//  LinReg.hpp
+//
+//  Created by Marc Melikyan on 10/2/20.
+//
+
+#include "core/math/math_defs.h"
+
+#include <string>
+#include <vector>
+
+class MLPPLinRegOld {
+public:
+	MLPPLinRegOld(std::vector<std::vector<real_t>> inputSet, std::vector<real_t> outputSet, std::string reg = "None", real_t lambda = 0.5, real_t alpha = 0.5);
+	std::vector<real_t> modelSetTest(std::vector<std::vector<real_t>> X);
+	real_t modelTest(std::vector<real_t> x);
+	void NewtonRaphson(real_t learning_rate, int max_epoch, bool UI);
+	void gradientDescent(real_t learning_rate, int max_epoch, bool UI = false);
+	void SGD(real_t learning_rate, int max_epoch, bool UI = false);
+
+	void Momentum(real_t learning_rate, int max_epoch, int mini_batch_size, real_t gamma, bool UI = false);
+	void NAG(real_t learning_rate, int max_epoch, int mini_batch_size, real_t gamma, bool UI = false);
+	void Adagrad(real_t learning_rate, int max_epoch, int mini_batch_size, real_t e, bool UI = false);
+	void Adadelta(real_t learning_rate, int max_epoch, int mini_batch_size, real_t b1, real_t e, bool UI = false);
+	void Adam(real_t learning_rate, int max_epoch, int mini_batch_size, real_t b1, real_t b2, real_t e, bool UI = false);
+	void Adamax(real_t learning_rate, int max_epoch, int mini_batch_size, real_t b1, real_t b2, real_t e, bool UI = false);
+	void Nadam(real_t learning_rate, int max_epoch, int mini_batch_size, real_t b1, real_t b2, real_t e, bool UI = false);
+
+	void MBGD(real_t learning_rate, int max_epoch, int mini_batch_size, bool UI = false);
+	void normalEquation();
+	real_t score();
+	void save(std::string fileName);
+
+private:
+	real_t Cost(std::vector<real_t> y_hat, std::vector<real_t> y);
+
+	std::vector<real_t> Evaluate(std::vector<std::vector<real_t>> X);
+	real_t Evaluate(std::vector<real_t> x);
+	void forwardPass();
+
+	std::vector<std::vector<real_t>> inputSet;
+	std::vector<real_t> outputSet;
+	std::vector<real_t> y_hat;
+	std::vector<real_t> weights;
+	real_t bias;
+
+	int n;
+	int k;
+
+	// Regularization Params
+	std::string reg;
+	int lambda;
+	int alpha; /* This is the controlling param for Elastic Net*/
+};
+
+#endif /* LinReg_hpp */

mlpp/log_reg/log_reg_old.cpp

@@ -0,0 +1,213 @@
+//
+//  LogReg.cpp
+//
+//  Created by Marc Melikyan on 10/2/20.
+//
+
+#include "log_reg_old.h"
+
+#include "../activation/activation.h"
+#include "../cost/cost.h"
+#include "../lin_alg/lin_alg.h"
+#include "../regularization/reg.h"
+#include "../utilities/utilities.h"
+
+#include <iostream>
+#include <random>
+
+MLPPLogRegOld::MLPPLogRegOld(std::vector<std::vector<real_t>> pinputSet, std::vector<real_t> poutputSet, std::string preg, real_t plambda, real_t palpha) {
+	inputSet = pinputSet;
+	outputSet = poutputSet;
+	n = pinputSet.size();
+	k = pinputSet[0].size();
+	reg = preg;
+	lambda = plambda;
+	alpha = palpha;
+
+	y_hat.resize(n);
+	weights = MLPPUtilities::weightInitialization(k);
+	bias = MLPPUtilities::biasInitialization();
+}
+
+std::vector<real_t> MLPPLogRegOld::modelSetTest(std::vector<std::vector<real_t>> X) {
+	return Evaluate(X);
+}
+
+real_t MLPPLogRegOld::modelTest(std::vector<real_t> x) {
+	return Evaluate(x);
+}
+
+void MLPPLogRegOld::gradientDescent(real_t learning_rate, int max_epoch, bool UI) {
+	MLPPLinAlg alg;
+	MLPPReg regularization;
+	real_t cost_prev = 0;
+	int epoch = 1;
+	forwardPass();
+
+	while (true) {
+		cost_prev = Cost(y_hat, outputSet);
+
+		std::vector<real_t> error = alg.subtraction(y_hat, outputSet);
+
+		// Calculating the weight gradients
+		weights = alg.subtraction(weights, alg.scalarMultiply(learning_rate / n, alg.mat_vec_mult(alg.transpose(inputSet), error)));
+		weights = regularization.regWeights(weights, lambda, alpha, reg);
+
+		// Calculating the bias gradients
+		bias -= learning_rate * alg.sum_elements(error) / n;
+		forwardPass();
+
+		if (UI) {
+			MLPPUtilities::CostInfo(epoch, cost_prev, Cost(y_hat, outputSet));
+			MLPPUtilities::UI(weights, bias);
+		}
+		epoch++;
+
+		if (epoch > max_epoch) {
+			break;
+		}
+	}
+}
+
+void MLPPLogRegOld::MLE(real_t learning_rate, int max_epoch, bool UI) {
+	MLPPLinAlg alg;
+	MLPPReg regularization;
+	real_t cost_prev = 0;
+	int epoch = 1;
+	forwardPass();
+
+	while (true) {
+		cost_prev = Cost(y_hat, outputSet);
+
+		std::vector<real_t> error = alg.subtraction(outputSet, y_hat);
+
+		// Calculating the weight gradients
+		weights = alg.addition(weights, alg.scalarMultiply(learning_rate / n, alg.mat_vec_mult(alg.transpose(inputSet), error)));
+		weights = regularization.regWeights(weights, lambda, alpha, reg);
+
+		// Calculating the bias gradients
+		bias += learning_rate * alg.sum_elements(error) / n;
+		forwardPass();
+
+		if (UI) {
+			MLPPUtilities::CostInfo(epoch, cost_prev, Cost(y_hat, outputSet));
+			MLPPUtilities::UI(weights, bias);
+		}
+		epoch++;
+		if (epoch > max_epoch) {
+			break;
+		}
+	}
+}
+
+void MLPPLogRegOld::SGD(real_t learning_rate, int max_epoch, bool UI) {
+	MLPPLinAlg alg;
+	MLPPReg regularization;
+	real_t cost_prev = 0;
+	int epoch = 1;
+
+	while (true) {
+		std::random_device rd;
+		std::default_random_engine generator(rd());
+		std::uniform_int_distribution<int> distribution(0, int(n - 1));
+		int outputIndex = distribution(generator);
+
+		real_t y_hat = Evaluate(inputSet[outputIndex]);
+		cost_prev = Cost({ y_hat }, { outputSet[outputIndex] });
+
+		real_t error = y_hat - outputSet[outputIndex];
+
+		// Weight updation
+		weights = alg.subtraction(weights, alg.scalarMultiply(learning_rate * error, inputSet[outputIndex]));
+		weights = regularization.regWeights(weights, lambda, alpha, reg);
+
+		// Bias updation
+		bias -= learning_rate * error;
+
+		y_hat = Evaluate({ inputSet[outputIndex] });
+
+		if (UI) {
+			MLPPUtilities::CostInfo(epoch, cost_prev, Cost({ y_hat }, { outputSet[outputIndex] }));
+			MLPPUtilities::UI(weights, bias);
+		}
+		epoch++;
+
+		if (epoch > max_epoch) {
+			break;
+		}
+	}
+	forwardPass();
+}
+
+void MLPPLogRegOld::MBGD(real_t learning_rate, int max_epoch, int mini_batch_size, bool UI) {
+	MLPPLinAlg alg;
+	MLPPReg regularization;
+	real_t cost_prev = 0;
+	int epoch = 1;
+
+	// Creating the mini-batches
+	int n_mini_batch = n / mini_batch_size;
+	auto bacthes = MLPPUtilities::createMiniBatches(inputSet, outputSet, n_mini_batch);
+	auto inputMiniBatches = std::get<0>(bacthes);
+	auto outputMiniBatches = std::get<1>(bacthes);
+
+	while (true) {
+		for (int i = 0; i < n_mini_batch; i++) {
+			std::vector<real_t> y_hat = Evaluate(inputMiniBatches[i]);
+			cost_prev = Cost(y_hat, outputMiniBatches[i]);
+
+			std::vector<real_t> error = alg.subtraction(y_hat, outputMiniBatches[i]);
+
+			// Calculating the weight gradients
+			weights = alg.subtraction(weights, alg.scalarMultiply(learning_rate / outputMiniBatches[i].size(), alg.mat_vec_mult(alg.transpose(inputMiniBatches[i]), error)));
+			weights = regularization.regWeights(weights, lambda, alpha, reg);
+
+			// Calculating the bias gradients
+			bias -= learning_rate * alg.sum_elements(error) / outputMiniBatches[i].size();
+			y_hat = Evaluate(inputMiniBatches[i]);
+
+			if (UI) {
+				MLPPUtilities::CostInfo(epoch, cost_prev, Cost(y_hat, outputMiniBatches[i]));
+				MLPPUtilities::UI(weights, bias);
+			}
+		}
+		epoch++;
+		if (epoch > max_epoch) {
+			break;
+		}
+	}
+	forwardPass();
+}
+
+real_t MLPPLogRegOld::score() {
+	MLPPUtilities util;
+	return util.performance(y_hat, outputSet);
+}
+
+void MLPPLogRegOld::save(std::string fileName) {
+	MLPPUtilities util;
+	util.saveParameters(fileName, weights, bias);
+}
+
+real_t MLPPLogRegOld::Cost(std::vector<real_t> y_hat, std::vector<real_t> y) {
+	MLPPReg regularization;
+	class MLPPCost cost;
+	return cost.LogLoss(y_hat, y) + regularization.regTerm(weights, lambda, alpha, reg);
+}
+
+std::vector<real_t> MLPPLogRegOld::Evaluate(std::vector<std::vector<real_t>> X) {
+	MLPPLinAlg alg;
+	MLPPActivation avn;
+	return avn.sigmoid(alg.scalarAdd(bias, alg.mat_vec_mult(X, weights)));
+}
+
+real_t MLPPLogRegOld::Evaluate(std::vector<real_t> x) {
+	MLPPLinAlg alg;
+	MLPPActivation avn;
+	return avn.sigmoid(alg.dot(weights, x) + bias);
+}
+
+// sigmoid ( wTx + b )
+void MLPPLogRegOld::forwardPass() {
+	y_hat = Evaluate(inputSet);
+}

mlpp/log_reg/log_reg_old.h

@@ -0,0 +1,51 @@
+
+#ifndef MLPP_LOG_REG_OLD_H
+#define MLPP_LOG_REG_OLD_H
+
+//
+//  LogReg.hpp
+//
+//  Created by Marc Melikyan on 10/2/20.
+//
+
+#include "core/math/math_defs.h"
+
+#include <string>
+#include <vector>
+
+class MLPPLogRegOld {
+public:
+	MLPPLogRegOld(std::vector<std::vector<real_t>> inputSet, std::vector<real_t> outputSet, std::string reg = "None", real_t lambda = 0.5, real_t alpha = 0.5);
+	std::vector<real_t> modelSetTest(std::vector<std::vector<real_t>> X);
+	real_t modelTest(std::vector<real_t> x);
+	void gradientDescent(real_t learning_rate, int max_epoch, bool UI = false);
+	void MLE(real_t learning_rate, int max_epoch, bool UI = false);
+	void SGD(real_t learning_rate, int max_epoch, bool UI = false);
+	void MBGD(real_t learning_rate, int max_epoch, int mini_batch_size, bool UI = false);
+	real_t score();
+	void save(std::string fileName);
+
+private:
+	real_t Cost(std::vector<real_t> y_hat, std::vector<real_t> y);
+
+	std::vector<real_t> Evaluate(std::vector<std::vector<real_t>> X);
+	real_t Evaluate(std::vector<real_t> x);
+	void forwardPass();
+
+	std::vector<std::vector<real_t>> inputSet;
+	std::vector<real_t> outputSet;
+	std::vector<real_t> y_hat;
+	std::vector<real_t> weights;
+	real_t bias;
+
+	int n;
+	int k;
+	real_t learning_rate;
+
+	// Regularization Params
+	std::string reg;
+	real_t lambda; /* Regularization Parameter */
+	real_t alpha; /* This is the controlling param for Elastic Net*/
+};
+
+#endif /* LogReg_hpp */

mlpp/mann/mann_old.cpp

@@ -0,0 +1,189 @@
+//
+//  MANN.cpp
+//
+//  Created by Marc Melikyan on 11/4/20.
+//
+
+#include "mann_old.h"
+
+#include "../activation/activation.h"
+#include "../cost/cost.h"
+#include "../lin_alg/lin_alg.h"
+#include "../regularization/reg.h"
+#include "../utilities/utilities.h"
+
+#include <iostream>
+
+MLPPMANNOld::MLPPMANNOld(std::vector<std::vector<real_t>> inputSet, std::vector<std::vector<real_t>> outputSet) :
+		inputSet(inputSet), outputSet(outputSet), n(inputSet.size()), k(inputSet[0].size()), n_output(outputSet[0].size()) {
+}
+
+MLPPMANNOld::~MLPPMANNOld() {
+	delete outputLayer;
+}
+
+std::vector<std::vector<real_t>> MLPPMANNOld::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;
+}
+
+std::vector<real_t> MLPPMANNOld::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 MLPPMANNOld::gradientDescent(real_t learning_rate, int max_epoch, bool UI) {
+	class MLPPCost cost;
+	MLPPActivation avn;
+	MLPPLinAlg alg;
+	MLPPReg regularization;
+
+	real_t cost_prev = 0;
+	int epoch = 1;
+	forwardPass();
+
+	while (true) {
+		cost_prev = Cost(y_hat, outputSet);
+
+		if (outputLayer->activation == "Softmax") {
+			outputLayer->delta = alg.subtraction(y_hat, outputSet);
+		} else {
+			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<std::vector<real_t>> outputWGrad = alg.matmult(alg.transpose(outputLayer->input), outputLayer->delta);
+
+		outputLayer->weights = alg.subtraction(outputLayer->weights, alg.scalarMultiply(learning_rate / n, outputWGrad));
+		outputLayer->weights = regularization.regWeights(outputLayer->weights, outputLayer->lambda, outputLayer->alpha, outputLayer->reg);
+		outputLayer->bias = alg.subtractMatrixRows(outputLayer->bias, alg.scalarMultiply(learning_rate / n, outputLayer->delta));
+
+		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.matmult(outputLayer->delta, alg.transpose(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);
+
+			network[network.size() - 1].weights = alg.subtraction(network[network.size() - 1].weights, alg.scalarMultiply(learning_rate / n, hiddenLayerWGrad));
+			network[network.size() - 1].weights = regularization.regWeights(network[network.size() - 1].weights, network[network.size() - 1].lambda, network[network.size() - 1].alpha, network[network.size() - 1].reg);
+			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--) {
+				hiddenLayerAvn = network[i].activation_map[network[i].activation];
+				network[i].delta = alg.hadamard_product(alg.matmult(network[i + 1].delta, network[i + 1].weights), (avn.*hiddenLayerAvn)(network[i].z, 1));
+				hiddenLayerWGrad = alg.matmult(alg.transpose(network[i].input), network[i].delta);
+				network[i].weights = alg.subtraction(network[i].weights, alg.scalarMultiply(learning_rate / n, hiddenLayerWGrad));
+				network[i].weights = regularization.regWeights(network[i].weights, network[i].lambda, network[i].alpha, network[i].reg);
+				network[i].bias = alg.subtractMatrixRows(network[i].bias, alg.scalarMultiply(learning_rate / n, network[i].delta));
+			}
+		}
+
+		forwardPass();
+
+		if (UI) {
+			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()) {
+				std::cout << "Layer " << network.size() << ": " << std::endl;
+				for (int i = network.size() - 1; i >= 0; i--) {
+					std::cout << "Layer " << i + 1 << ": " << std::endl;
+					MLPPUtilities::UI(network[i].weights, network[i].bias);
+				}
+			}
+		}
+
+		epoch++;
+		if (epoch > max_epoch) {
+			break;
+		}
+	}
+}
+
+real_t MLPPMANNOld::score() {
+	MLPPUtilities util;
+	forwardPass();
+	return util.performance(y_hat, outputSet);
+}
+
+void MLPPMANNOld::save(std::string fileName) {
+	MLPPUtilities util;
+	if (!network.empty()) {
+		util.saveParameters(fileName, network[0].weights, network[0].bias, 0, 1);
+		for (uint32_t 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);
+	}
+}
+
+void MLPPMANNOld::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 MLPPMANNOld::addOutputLayer(std::string activation, std::string loss, std::string weightInit, std::string reg, real_t lambda, real_t alpha) {
+	if (!network.empty()) {
+		outputLayer = new MLPPOldMultiOutputLayer(n_output, network[0].n_hidden, activation, loss, network[network.size() - 1].a, weightInit, reg, lambda, alpha);
+	} else {
+		outputLayer = new MLPPOldMultiOutputLayer(n_output, k, activation, loss, inputSet, weightInit, reg, lambda, alpha);
+	}
+}
+
+real_t MLPPMANNOld::Cost(std::vector<std::vector<real_t>> y_hat, std::vector<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 MLPPMANNOld::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;
+}

mlpp/mann/mann_old.h

@@ -0,0 +1,51 @@
+
+#ifndef MLPP_MANN_OLD_H
+#define MLPP_MANN_OLD_H
+
+//
+//  MANN.hpp
+//
+//  Created by Marc Melikyan on 11/4/20.
+//
+
+#include "core/math/math_defs.h"
+
+#include "../hidden_layer/hidden_layer.h"
+#include "../multi_output_layer/multi_output_layer.h"
+
+#include "../hidden_layer/hidden_layer_old.h"
+#include "../multi_output_layer/multi_output_layer_old.h"
+
+#include <string>
+#include <vector>
+
+class MLPPMANNOld {
+public:
+	MLPPMANNOld(std::vector<std::vector<real_t>> inputSet, std::vector<std::vector<real_t>> outputSet);
+	~MLPPMANNOld();
+	std::vector<std::vector<real_t>> modelSetTest(std::vector<std::vector<real_t>> X);
+	std::vector<real_t> modelTest(std::vector<real_t> x);
+	void gradientDescent(real_t learning_rate, int max_epoch, bool UI = false);
+	real_t score();
+	void save(std::string fileName);
+
+	void addLayer(int n_hidden, std::string activation, std::string weightInit = "Default", std::string reg = "None", real_t lambda = 0.5, real_t alpha = 0.5);
+	void addOutputLayer(std::string activation, std::string loss, std::string weightInit = "Default", std::string reg = "None", real_t lambda = 0.5, real_t alpha = 0.5);
+
+private:
+	real_t Cost(std::vector<std::vector<real_t>> y_hat, std::vector<std::vector<real_t>> y);
+	void forwardPass();
+
+	std::vector<std::vector<real_t>> inputSet;
+	std::vector<std::vector<real_t>> outputSet;
+	std::vector<std::vector<real_t>> y_hat;
+
+	std::vector<MLPPOldHiddenLayer> network;
+	MLPPOldMultiOutputLayer *outputLayer;
+
+	int n;
+	int k;
+	int n_output;
+};
+
+#endif /* MANN_hpp */

mlpp/multinomial_nb/multinomial_nb_old.cpp

@@ -0,0 +1,121 @@
+//
+//  MultinomialNB.cpp
+//
+//  Created by Marc Melikyan on 1/17/21.
+//
+
+#include "multinomial_nb_old.h"
+
+#include "../lin_alg/lin_alg.h"
+#include "../utilities/utilities.h"
+
+#include <algorithm>
+#include <iostream>
+#include <random>
+
+MLPPMultinomialNBOld::MLPPMultinomialNBOld(std::vector<std::vector<real_t>> pinputSet, std::vector<real_t> poutputSet, int pclass_num) {
+	inputSet = pinputSet;
+	outputSet = poutputSet;
+	class_num = pclass_num;
+
+	y_hat.resize(outputSet.size());
+	Evaluate();
+}
+
+std::vector<real_t> MLPPMultinomialNBOld::modelSetTest(std::vector<std::vector<real_t>> X) {
+	std::vector<real_t> y_hat;
+	for (uint32_t i = 0; i < X.size(); i++) {
+		y_hat.push_back(modelTest(X[i]));
+	}
+	return y_hat;
+}
+
+real_t MLPPMultinomialNBOld::modelTest(std::vector<real_t> x) {
+	real_t score[class_num];
+	computeTheta();
+
+	for (uint32_t j = 0; j < x.size(); j++) {
+		for (uint32_t k = 0; k < vocab.size(); k++) {
+			if (x[j] == vocab[k]) {
+				for (int p = class_num - 1; p >= 0; p--) {
+					score[p] += std::log(theta[p][vocab[k]]);
+				}
+			}
+		}
+	}
+
+	for (uint32_t i = 0; i < priors.size(); i++) {
+		score[i] += std::log(priors[i]);
+	}
+
+	return std::distance(score, std::max_element(score, score + sizeof(score) / sizeof(real_t)));
+}
+
+real_t MLPPMultinomialNBOld::score() {
+	MLPPUtilities util;
+	return util.performance(y_hat, outputSet);
+}
+
+void MLPPMultinomialNBOld::computeTheta() {
+	// Resizing theta for the sake of ease & proper access of the elements.
+	theta.resize(class_num);
+
+	// Setting all values in the hasmap by default to 0.
+	for (int i = class_num - 1; i >= 0; i--) {
+		for (uint32_t j = 0; j < vocab.size(); j++) {
+			theta[i][vocab[j]] = 0;
+		}
+	}
+
+	for (uint32_t i = 0; i < inputSet.size(); i++) {
+		for (uint32_t j = 0; j < inputSet[0].size(); j++) {
+			theta[outputSet[i]][inputSet[i][j]]++;
+		}
+	}
+
+	for (uint32_t i = 0; i < theta.size(); i++) {
+		for (uint32_t j = 0; j < theta[i].size(); j++) {
+			theta[i][j] /= priors[i] * y_hat.size();
+		}
+	}
+}
+
+void MLPPMultinomialNBOld::Evaluate() {
+	MLPPLinAlg alg;
+	for (uint32_t i = 0; i < outputSet.size(); i++) {
+		// Pr(B | A) * Pr(A)
+		real_t score[class_num];
+
+		// Easy computation of priors, i.e. Pr(C_k)
+		priors.resize(class_num);
+		for (uint32_t ii = 0; ii < outputSet.size(); ii++) {
+			priors[int(outputSet[ii])]++;
+		}
+		priors = alg.scalarMultiply(real_t(1) / real_t(outputSet.size()), priors);
+
+		// Evaluating Theta...
+		computeTheta();
+
+		for (uint32_t j = 0; j < inputSet.size(); j++) {
+			for (uint32_t k = 0; k < vocab.size(); k++) {
+				if (inputSet[i][j] == vocab[k]) {
+					for (int p = class_num - 1; p >= 0; p--) {
+						score[p] += std::log(theta[i][vocab[k]]);
+					}
+				}
+			}
+		}
+
+		for (uint32_t ii = 0; ii < priors.size(); ii++) {
+			score[ii] += std::log(priors[ii]);
+			score[ii] = exp(score[ii]);
+		}
+
+		for (int ii = 0; ii < 2; ii++) {
+			std::cout << score[ii] << std::endl;
+		}
+
+		// Assigning the traning example's y_hat to a class
+		y_hat[i] = std::distance(score, std::max_element(score, score + sizeof(score) / sizeof(real_t)));
+	}
+}

mlpp/multinomial_nb/multinomial_nb_old.h

@@ -0,0 +1,40 @@
+
+#ifndef MLPP_MULTINOMIAL_NB_OLD_H
+#define MLPP_MULTINOMIAL_NB_OLD_H
+
+//
+//  MultinomialNB.hpp
+//
+//  Created by Marc Melikyan on 1/17/21.
+//
+
+#include "core/math/math_defs.h"
+
+#include <map>
+#include <vector>
+
+class MLPPMultinomialNBOld {
+public:
+	MLPPMultinomialNBOld(std::vector<std::vector<real_t>> inputSet, std::vector<real_t> outputSet, int class_num);
+	std::vector<real_t> modelSetTest(std::vector<std::vector<real_t>> X);
+	real_t modelTest(std::vector<real_t> x);
+	real_t score();
+
+private:
+	void computeTheta();
+	void Evaluate();
+
+	// Model Params
+	std::vector<real_t> priors;
+
+	std::vector<std::map<real_t, int>> theta;
+	std::vector<real_t> vocab;
+	int class_num;
+
+	// Datasets
+	std::vector<std::vector<real_t>> inputSet;
+	std::vector<real_t> outputSet;
+	std::vector<real_t> y_hat;
+};
+
+#endif /* MultinomialNB_hpp */

mlpp/softmax_net/softmax_net_old.cpp

@@ -0,0 +1,309 @@
+//
+//  SoftmaxNet.cpp
+//
+//  Created by Marc Melikyan on 10/2/20.
+//
+
+#include "softmax_net_old.h"
+
+#include "../activation/activation.h"
+#include "../cost/cost.h"
+#include "../data/data.h"
+#include "../lin_alg/lin_alg.h"
+#include "../regularization/reg.h"
+#include "../utilities/utilities.h"
+
+#include <iostream>
+#include <random>
+
+MLPPSoftmaxNetOld::MLPPSoftmaxNetOld(std::vector<std::vector<real_t>> pinputSet, std::vector<std::vector<real_t>> poutputSet, int pn_hidden, std::string preg, real_t plambda, real_t palpha) {
+	inputSet = pinputSet;
+	outputSet = poutputSet;
+	n = pinputSet.size();
+	k = pinputSet[0].size();
+	n_hidden = pn_hidden;
+	n_class = poutputSet[0].size();
+	reg = preg;
+	lambda = plambda;
+	alpha = palpha;
+
+	y_hat.resize(n);
+
+	weights1 = MLPPUtilities::weightInitialization(k, n_hidden);
+	weights2 = MLPPUtilities::weightInitialization(n_hidden, n_class);
+	bias1 = MLPPUtilities::biasInitialization(n_hidden);
+	bias2 = MLPPUtilities::biasInitialization(n_class);
+}
+
+std::vector<real_t> MLPPSoftmaxNetOld::modelTest(std::vector<real_t> x) {
+	return Evaluate(x);
+}
+
+std::vector<std::vector<real_t>> MLPPSoftmaxNetOld::modelSetTest(std::vector<std::vector<real_t>> X) {
+	return Evaluate(X);
+}
+
+void MLPPSoftmaxNetOld::gradientDescent(real_t learning_rate, int max_epoch, bool UI) {
+	MLPPActivation avn;
+	MLPPLinAlg alg;
+	MLPPReg regularization;
+	real_t cost_prev = 0;
+	int epoch = 1;
+	forwardPass();
+
+	while (true) {
+		cost_prev = Cost(y_hat, outputSet);
+
+		// Calculating the errors
+		std::vector<std::vector<real_t>> error = alg.subtraction(y_hat, outputSet);
+
+		// Calculating the weight/bias gradients for layer 2
+
+		std::vector<std::vector<real_t>> D2_1 = alg.matmult(alg.transpose(a2), error);
+
+		// weights and bias updation for layer 2
+		weights2 = alg.subtraction(weights2, alg.scalarMultiply(learning_rate, D2_1));
+		weights2 = regularization.regWeights(weights2, lambda, alpha, reg);
+
+		bias2 = alg.subtractMatrixRows(bias2, alg.scalarMultiply(learning_rate, error));
+
+		//Calculating the weight/bias for layer 1
+
+		std::vector<std::vector<real_t>> D1_1 = alg.matmult(error, alg.transpose(weights2));
+
+		std::vector<std::vector<real_t>> D1_2 = alg.hadamard_product(D1_1, avn.sigmoid(z2, 1));
+
+		std::vector<std::vector<real_t>> D1_3 = alg.matmult(alg.transpose(inputSet), D1_2);
+
+		// weight an bias updation for layer 1
+		weights1 = alg.subtraction(weights1, alg.scalarMultiply(learning_rate, D1_3));
+		weights1 = regularization.regWeights(weights1, lambda, alpha, reg);
+
+		bias1 = alg.subtractMatrixRows(bias1, alg.scalarMultiply(learning_rate, D1_2));
+
+		forwardPass();
+
+		// UI PORTION
+		if (UI) {
+			MLPPUtilities::CostInfo(epoch, cost_prev, Cost(y_hat, outputSet));
+			std::cout << "Layer 1:" << std::endl;
+			MLPPUtilities::UI(weights1, bias1);
+			std::cout << "Layer 2:" << std::endl;
+			MLPPUtilities::UI(weights2, bias2);
+		}
+		epoch++;
+
+		if (epoch > max_epoch) {
+			break;
+		}
+	}
+}
+
+void MLPPSoftmaxNetOld::SGD(real_t learning_rate, int max_epoch, bool UI) {
+	MLPPActivation avn;
+	MLPPLinAlg alg;
+	MLPPReg regularization;
+	real_t cost_prev = 0;
+	int epoch = 1;
+
+	while (true) {
+		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 = Evaluate(inputSet[outputIndex]);
+
+		auto prop_res = propagate(inputSet[outputIndex]);
+		auto z2 = std::get<0>(prop_res);
+		auto a2 = std::get<1>(prop_res);
+
+		cost_prev = Cost({ y_hat }, { outputSet[outputIndex] });
+		std::vector<real_t> error = alg.subtraction(y_hat, outputSet[outputIndex]);
+
+		// Weight updation for layer 2
+		std::vector<std::vector<real_t>> D2_1 = alg.outerProduct(error, a2);
+		weights2 = alg.subtraction(weights2, alg.scalarMultiply(learning_rate, alg.transpose(D2_1)));
+		weights2 = regularization.regWeights(weights2, lambda, alpha, reg);
+
+		// Bias updation for layer 2
+		bias2 = alg.subtraction(bias2, alg.scalarMultiply(learning_rate, error));
+
+		// Weight updation for layer 1
+		std::vector<real_t> D1_1 = alg.mat_vec_mult(weights2, error);
+		std::vector<real_t> D1_2 = alg.hadamard_product(D1_1, avn.sigmoid(z2, true));
+		std::vector<std::vector<real_t>> D1_3 = alg.outerProduct(inputSet[outputIndex], D1_2);
+
+		weights1 = alg.subtraction(weights1, alg.scalarMultiply(learning_rate, D1_3));
+		weights1 = regularization.regWeights(weights1, lambda, alpha, reg);
+		// Bias updation for layer 1
+
+		bias1 = alg.subtraction(bias1, alg.scalarMultiply(learning_rate, D1_2));
+
+		y_hat = Evaluate(inputSet[outputIndex]);
+		if (UI) {
+			MLPPUtilities::CostInfo(epoch, cost_prev, Cost({ y_hat }, { outputSet[outputIndex] }));
+			std::cout << "Layer 1:" << std::endl;
+			MLPPUtilities::UI(weights1, bias1);
+			std::cout << "Layer 2:" << std::endl;
+			MLPPUtilities::UI(weights2, bias2);
+		}
+		epoch++;
+
+		if (epoch > max_epoch) {
+			break;
+		}
+	}
+	forwardPass();
+}
+
+void MLPPSoftmaxNetOld::MBGD(real_t learning_rate, int max_epoch, int mini_batch_size, bool UI) {
+	MLPPActivation avn;
+	MLPPLinAlg alg;
+	MLPPReg regularization;
+	real_t cost_prev = 0;
+	int epoch = 1;
+
+	// Creating the mini-batches
+	int n_mini_batch = n / mini_batch_size;
+
+	auto batches = MLPPUtilities::createMiniBatches(inputSet, outputSet, n_mini_batch);
+	auto inputMiniBatches = std::get<0>(batches);
+	auto outputMiniBatches = std::get<1>(batches);
+
+	// Creating the mini-batches
+	for (int i = 0; i < n_mini_batch; i++) {
+		std::vector<std::vector<real_t>> currentInputSet;
+		std::vector<std::vector<real_t>> currentOutputSet;
+		for (int j = 0; j < n / n_mini_batch; j++) {
+			currentInputSet.push_back(inputSet[n / n_mini_batch * i + j]);
+			currentOutputSet.push_back(outputSet[n / n_mini_batch * i + j]);
+		}
+		inputMiniBatches.push_back(currentInputSet);
+		outputMiniBatches.push_back(currentOutputSet);
+	}
+
+	if (real_t(n) / real_t(n_mini_batch) - int(n / n_mini_batch) != 0) {
+		for (int i = 0; i < n - n / n_mini_batch * n_mini_batch; i++) {
+			inputMiniBatches[n_mini_batch - 1].push_back(inputSet[n / n_mini_batch * n_mini_batch + i]);
+			outputMiniBatches[n_mini_batch - 1].push_back(outputSet[n / n_mini_batch * n_mini_batch + i]);
+		}
+	}
+
+	while (true) {
+		for (int i = 0; i < n_mini_batch; i++) {
+			std::vector<std::vector<real_t>> y_hat = Evaluate(inputMiniBatches[i]);
+
+			auto propagate_res = propagate(inputMiniBatches[i]);
+			auto z2 = std::get<0>(propagate_res);
+			auto a2 = std::get<1>(propagate_res);
+
+			cost_prev = Cost(y_hat, outputMiniBatches[i]);
+
+			// Calculating the errors
+			std::vector<std::vector<real_t>> error = alg.subtraction(y_hat, outputMiniBatches[i]);
+
+			// Calculating the weight/bias gradients for layer 2
+
+			std::vector<std::vector<real_t>> D2_1 = alg.matmult(alg.transpose(a2), error);
+
+			// weights and bias updation for layser 2
+			weights2 = alg.subtraction(weights2, alg.scalarMultiply(learning_rate, D2_1));
+			weights2 = regularization.regWeights(weights2, lambda, alpha, reg);
+
+			// Bias Updation for layer 2
+			bias2 = alg.subtractMatrixRows(bias2, alg.scalarMultiply(learning_rate, error));
+
+			//Calculating the weight/bias for layer 1
+
+			std::vector<std::vector<real_t>> D1_1 = alg.matmult(error, alg.transpose(weights2));
+
+			std::vector<std::vector<real_t>> D1_2 = alg.hadamard_product(D1_1, avn.sigmoid(z2, 1));
+
+			std::vector<std::vector<real_t>> D1_3 = alg.matmult(alg.transpose(inputMiniBatches[i]), D1_2);
+
+			// weight an bias updation for layer 1
+			weights1 = alg.subtraction(weights1, alg.scalarMultiply(learning_rate, D1_3));
+			weights1 = regularization.regWeights(weights1, lambda, alpha, reg);
+
+			bias1 = alg.subtractMatrixRows(bias1, alg.scalarMultiply(learning_rate, D1_2));
+
+			y_hat = Evaluate(inputMiniBatches[i]);
+
+			if (UI) {
+				MLPPUtilities::CostInfo(epoch, cost_prev, Cost(y_hat, outputMiniBatches[i]));
+				std::cout << "Layer 1:" << std::endl;
+				MLPPUtilities::UI(weights1, bias1);
+				std::cout << "Layer 2:" << std::endl;
+				MLPPUtilities::UI(weights2, bias2);
+			}
+		}
+		epoch++;
+		if (epoch > max_epoch) {
+			break;
+		}
+	}
+	forwardPass();
+}
+
+real_t MLPPSoftmaxNetOld::score() {
+	MLPPUtilities util;
+	return util.performance(y_hat, outputSet);
+}
+
+void MLPPSoftmaxNetOld::save(std::string fileName) {
+	MLPPUtilities util;
+	util.saveParameters(fileName, weights1, bias1, 0, 1);
+	util.saveParameters(fileName, weights2, bias2, 1, 2);
+}
+
+std::vector<std::vector<real_t>> MLPPSoftmaxNetOld::getEmbeddings() {
+	return weights1;
+}
+
+real_t MLPPSoftmaxNetOld::Cost(std::vector<std::vector<real_t>> y_hat, std::vector<std::vector<real_t>> y) {
+	MLPPReg regularization;
+	MLPPData data;
+	class MLPPCost cost;
+	return cost.CrossEntropy(y_hat, y) + regularization.regTerm(weights1, lambda, alpha, reg) + regularization.regTerm(weights2, lambda, alpha, reg);
+}
+
+std::vector<std::vector<real_t>> MLPPSoftmaxNetOld::Evaluate(std::vector<std::vector<real_t>> X) {
+	MLPPLinAlg alg;
+	MLPPActivation avn;
+	std::vector<std::vector<real_t>> z2 = alg.mat_vec_add(alg.matmult(X, weights1), bias1);
+	std::vector<std::vector<real_t>> a2 = avn.sigmoid(z2);
+	return avn.adjSoftmax(alg.mat_vec_add(alg.matmult(a2, weights2), bias2));
+}
+
+std::tuple<std::vector<std::vector<real_t>>, std::vector<std::vector<real_t>>> MLPPSoftmaxNetOld::propagate(std::vector<std::vector<real_t>> X) {
+	MLPPLinAlg alg;
+	MLPPActivation avn;
+	std::vector<std::vector<real_t>> z2 = alg.mat_vec_add(alg.matmult(X, weights1), bias1);
+	std::vector<std::vector<real_t>> a2 = avn.sigmoid(z2);
+	return { z2, a2 };
+}
+
+std::vector<real_t> MLPPSoftmaxNetOld::Evaluate(std::vector<real_t> x) {
+	MLPPLinAlg alg;
+	MLPPActivation avn;
+	std::vector<real_t> z2 = alg.addition(alg.mat_vec_mult(alg.transpose(weights1), x), bias1);
+	std::vector<real_t> a2 = avn.sigmoid(z2);
+	return avn.adjSoftmax(alg.addition(alg.mat_vec_mult(alg.transpose(weights2), a2), bias2));
+}
+
+std::tuple<std::vector<real_t>, std::vector<real_t>> MLPPSoftmaxNetOld::propagate(std::vector<real_t> x) {
+	MLPPLinAlg alg;
+	MLPPActivation avn;
+	std::vector<real_t> z2 = alg.addition(alg.mat_vec_mult(alg.transpose(weights1), x), bias1);
+	std::vector<real_t> a2 = avn.sigmoid(z2);
+	return { z2, a2 };
+}
+
+void MLPPSoftmaxNetOld::forwardPass() {
+	MLPPLinAlg alg;
+	MLPPActivation avn;
+	z2 = alg.mat_vec_add(alg.matmult(inputSet, weights1), bias1);
+	a2 = avn.sigmoid(z2);
+	y_hat = avn.adjSoftmax(alg.mat_vec_add(alg.matmult(a2, weights2), bias2));
+}

mlpp/softmax_net/softmax_net_old.h

@@ -0,0 +1,60 @@
+#ifndef MLPP_SOFTMAX_NET_OLD_H
+#define MLPP_SOFTMAX_NET_OLD_H
+
+//
+//  SoftmaxNet.hpp
+//
+//  Created by Marc Melikyan on 10/2/20.
+//
+
+#include "core/math/math_defs.h"
+
+#include <string>
+#include <vector>
+
+class MLPPSoftmaxNetOld {
+public:
+	MLPPSoftmaxNetOld(std::vector<std::vector<real_t>> inputSet, std::vector<std::vector<real_t>> outputSet, int n_hidden, std::string reg = "None", real_t lambda = 0.5, real_t alpha = 0.5);
+	std::vector<real_t> modelTest(std::vector<real_t> x);
+	std::vector<std::vector<real_t>> modelSetTest(std::vector<std::vector<real_t>> X);
+	void gradientDescent(real_t learning_rate, int max_epoch, bool UI = false);
+	void SGD(real_t learning_rate, int max_epoch, bool UI = false);
+	void MBGD(real_t learning_rate, int max_epoch, int mini_batch_size, bool UI = false);
+	real_t score();
+	void save(std::string fileName);
+
+	std::vector<std::vector<real_t>> getEmbeddings(); // This class is used (mostly) for word2Vec. This function returns our embeddings.
+private:
+	real_t Cost(std::vector<std::vector<real_t>> y_hat, std::vector<std::vector<real_t>> y);
+
+	std::vector<std::vector<real_t>> Evaluate(std::vector<std::vector<real_t>> X);
+	std::tuple<std::vector<std::vector<real_t>>, std::vector<std::vector<real_t>>> propagate(std::vector<std::vector<real_t>> X);
+	std::vector<real_t> Evaluate(std::vector<real_t> x);
+	std::tuple<std::vector<real_t>, std::vector<real_t>> propagate(std::vector<real_t> x);
+	void forwardPass();
+
+	std::vector<std::vector<real_t>> inputSet;
+	std::vector<std::vector<real_t>> outputSet;
+	std::vector<std::vector<real_t>> y_hat;
+
+	std::vector<std::vector<real_t>> weights1;
+	std::vector<std::vector<real_t>> weights2;
+
+	std::vector<real_t> bias1;
+	std::vector<real_t> bias2;
+
+	std::vector<std::vector<real_t>> z2;
+	std::vector<std::vector<real_t>> a2;
+
+	int n;
+	int k;
+	int n_class;
+	int n_hidden;
+
+	// Regularization Params
+	std::string reg;
+	real_t lambda;
+	real_t alpha; /* This is the controlling param for Elastic Net*/
+};
+
+#endif /* SoftmaxNet_hpp */