pmlpp/mlpp/wgan/wgan.cpp

//
//  WGAN.cpp
//
//  Created by Marc Melikyan on 11/4/20.
//

#include "wgan.h"

#include "core/log/logger.h"

#include "../activation/activation.h"
#include "../cost/cost.h"
#include "../lin_alg/lin_alg.h"
#include "../regularization/reg.h"
#include "../utilities/utilities.h"

#include "core/object/method_bind_ext.gen.inc"

Ref<MLPPMatrix> MLPPWGAN::get_output_set() {
	return output_set;
}
void MLPPWGAN::set_output_set(const Ref<MLPPMatrix> &val) {
	output_set = val;

	n = 0;

	if (output_set.is_valid()) {
		n = output_set->size().y;
	}
}

int MLPPWGAN::get_k() const {
	return k;
}
void MLPPWGAN::set_k(const int val) {
	k = val;
}

Ref<MLPPMatrix> MLPPWGAN::generate_example(int n) {
	MLPPLinAlg alg;

	return model_set_test_generator(alg.gaussian_noise(n, k));
}

void MLPPWGAN::gradient_descent(real_t learning_rate, int max_epoch, bool ui) {
	//MLPPCost mlpp_cost;
	MLPPLinAlg alg;
	real_t cost_prev = 0;
	int epoch = 1;

	forward_pass();

	const int CRITIC_INTERATIONS = 5; // Wasserstein GAN specific parameter.

	while (true) {
		cost_prev = cost(y_hat, alg.onevecv(n));

		Ref<MLPPMatrix> generator_input_set;
		Ref<MLPPMatrix> discriminator_input_set;
		discriminator_input_set.instance();

		Ref<MLPPVector> ly_hat;
		Ref<MLPPVector> loutput_set;

		// Training of the discriminator.
		for (int i = 0; i < CRITIC_INTERATIONS; i++) {
			generator_input_set = alg.gaussian_noise(n, k);
			discriminator_input_set->set_from_mlpp_matrix(model_set_test_generator(generator_input_set));
			discriminator_input_set->add_rows_mlpp_matrix(output_set); // Fake + real inputs.

			ly_hat = model_set_test_discriminator(discriminator_input_set);
			loutput_set = alg.scalar_multiplynv(-1, alg.onevecv(n)); // WGAN changes y_i = 1 and y_i = 0 to y_i = 1 and y_i = -1
			Ref<MLPPVector> output_set_real = alg.onevecv(n);
			loutput_set->add_mlpp_vector(output_set_real); // Fake + real output scores.

			DiscriminatorGradientResult discriminator_gradient_results = compute_discriminator_gradients(ly_hat, loutput_set);
			Vector<Ref<MLPPMatrix>> cumulative_discriminator_hidden_layer_w_grad = discriminator_gradient_results.cumulative_hidden_layer_w_grad;
			Ref<MLPPVector> output_discriminator_w_grad = discriminator_gradient_results.output_w_grad;

			cumulative_discriminator_hidden_layer_w_grad = alg.scalar_multiply_vm(learning_rate / n, cumulative_discriminator_hidden_layer_w_grad);
			output_discriminator_w_grad = alg.scalar_multiplynv(learning_rate / n, output_discriminator_w_grad);
			update_discriminator_parameters(cumulative_discriminator_hidden_layer_w_grad, output_discriminator_w_grad, learning_rate);
		}

		// Training of the generator.
		generator_input_set = alg.gaussian_noise(n, k);
		discriminator_input_set->set_from_mlpp_matrix(model_set_test_generator(generator_input_set));
		ly_hat = model_set_test_discriminator(discriminator_input_set);
		loutput_set = alg.onevecv(n);

		Vector<Ref<MLPPMatrix>> cumulative_generator_hidden_layer_w_grad = compute_generator_gradients(y_hat, loutput_set);
		cumulative_generator_hidden_layer_w_grad = alg.scalar_multiply_vm(learning_rate / n, cumulative_generator_hidden_layer_w_grad);
		update_generator_parameters(cumulative_generator_hidden_layer_w_grad, learning_rate);

		forward_pass();

		if (ui) {
			handle_ui(epoch, cost_prev, y_hat, alg.onevecv(n));
		}

		epoch++;
		if (epoch > max_epoch) {
			break;
		}
	}
}

real_t MLPPWGAN::score() {
	MLPPLinAlg alg;
	MLPPUtilities util;
	forward_pass();
	return util.performance_vec(y_hat, alg.onevecv(n));
}

void MLPPWGAN::save(const String &file_name) {
	MLPPUtilities util;

	/*
	if (!network.empty()) {
		util.saveParameters(file_name, 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(file_name, outputLayer->weights, outputLayer->bias, 1, network.size() + 1);
	} else {
		util.saveParameters(file_name, outputLayer->weights, outputLayer->bias, 0, network.size() + 1);
	}
	*/
}

void MLPPWGAN::add_layer(int n_hidden, MLPPActivation::ActivationFunction activation, MLPPUtilities::WeightDistributionType weight_init, MLPPReg::RegularizationType reg, real_t lambda, real_t alpha) {
	MLPPLinAlg alg;

	Ref<MLPPHiddenLayer> layer;
	layer.instance();

	layer->set_n_hidden(n_hidden);
	layer->set_activation(activation);
	layer->set_weight_init(weight_init);
	layer->set_reg(reg);
	layer->set_lambda(lambda);
	layer->set_alpha(alpha);

	if (network.empty()) {
		layer->set_input(alg.gaussian_noise(n, k));
	} else {
		layer->set_input(network.write[network.size() - 1]->get_a());
	}

	network.push_back(layer);
	layer->forward_pass();
}

void MLPPWGAN::add_output_layer(MLPPUtilities::WeightDistributionType weight_init, MLPPReg::RegularizationType reg, real_t lambda, real_t alpha) {
	ERR_FAIL_COND(network.empty());

	if (!output_layer.is_valid()) {
		output_layer.instance();
	}

	output_layer->set_n_hidden(network[network.size() - 1]->get_n_hidden());
	output_layer->set_activation(MLPPActivation::ACTIVATION_FUNCTION_LINEAR);
	output_layer->set_cost(MLPPCost::COST_TYPE_WASSERSTEIN_LOSS);
	output_layer->set_input(network.write[network.size() - 1]->get_a());
	output_layer->set_weight_init(weight_init);
	output_layer->set_lambda(lambda);
	output_layer->set_alpha(alpha);
}

MLPPWGAN::MLPPWGAN(real_t p_k, const Ref<MLPPMatrix> &p_output_set) {
	output_set = p_output_set;
	n = p_output_set->size().y;
	k = p_k;

	y_hat.instance();
}

MLPPWGAN::MLPPWGAN() {
	n = 0;
	k = 0;

	y_hat.instance();
}

MLPPWGAN::~MLPPWGAN() {
}

Ref<MLPPMatrix> MLPPWGAN::model_set_test_generator(const Ref<MLPPMatrix> &X) {
	if (!network.empty()) {
		network.write[0]->set_input(X);
		network.write[0]->forward_pass();

		for (int i = 1; i <= network.size() / 2; ++i) {
			network.write[i]->set_input(network.write[i - 1]->get_a());
			network.write[i]->forward_pass();
		}
	}

	return network.write[network.size() / 2]->get_a();
}

Ref<MLPPVector> MLPPWGAN::model_set_test_discriminator(const Ref<MLPPMatrix> &X) {
	if (!network.empty()) {
		for (int i = network.size() / 2 + 1; i < network.size(); i++) {
			if (i == network.size() / 2 + 1) {
				network.write[i]->set_input(X);
			} else {
				network.write[i]->set_input(network.write[i - 1]->get_a());
			}
			network.write[i]->forward_pass();
		}

		output_layer->set_input(network.write[network.size() - 1]->get_a());
	}

	output_layer->forward_pass();

	return output_layer->get_a();
}

real_t MLPPWGAN::cost(const Ref<MLPPVector> &y_hat, const Ref<MLPPVector> &y) {
	MLPPReg regularization;
	MLPPCost mlpp_cost;

	real_t total_reg_term = 0;

	for (int i = 0; i < network.size() - 1; ++i) {
		Ref<MLPPHiddenLayer> layer = network[i];

		total_reg_term += regularization.reg_termm(layer->get_weights(), layer->get_lambda(), layer->get_alpha(), layer->get_reg());
	}

	total_reg_term += regularization.reg_termv(output_layer->get_weights(), output_layer->get_lambda(), output_layer->get_alpha(), output_layer->get_reg());

	return mlpp_cost.run_cost_norm_vector(output_layer->get_cost(), y_hat, y) + total_reg_term;
}

void MLPPWGAN::forward_pass() {
	MLPPLinAlg alg;

	if (!network.empty()) {
		Ref<MLPPHiddenLayer> layer = network[0];

		layer->set_input(alg.gaussian_noise(n, k));
		layer->forward_pass();

		for (int i = 1; i < network.size(); i++) {
			layer = network[i];

			layer->set_input(network.write[i - 1]->get_a());
			layer->forward_pass();
		}

		output_layer->set_input(network.write[network.size() - 1]->get_a());
	} else { // Should never happen, though.
		output_layer->set_input(alg.gaussian_noise(n, k));
	}

	output_layer->forward_pass();

	y_hat->set_from_mlpp_vector(output_layer->get_a());
}

void MLPPWGAN::update_discriminator_parameters(Vector<Ref<MLPPMatrix>> hidden_layer_updations, const Ref<MLPPVector> &output_layer_updation, real_t learning_rate) {
	MLPPLinAlg alg;

	output_layer->set_weights(alg.subtractionnv(output_layer->get_weights(), output_layer_updation));
	output_layer->set_bias(output_layer->get_bias() - learning_rate * alg.sum_elementsv(output_layer->get_delta()) / n);

	if (!network.empty()) {
		Ref<MLPPHiddenLayer> layer = network[network.size() - 1];

		layer->set_weights(alg.subtractionm(layer->get_weights(), hidden_layer_updations[0]));
		layer->set_bias(alg.subtract_matrix_rows(layer->get_bias(), alg.scalar_multiplym(learning_rate / n, layer->get_delta())));

		for (int i = network.size() - 2; i > network.size() / 2; i--) {
			layer = network[i];

			layer->set_weights(alg.subtractionm(layer->get_weights(), hidden_layer_updations[(network.size() - 2) - i + 1]));
			layer->set_bias(alg.subtract_matrix_rows(layer->get_bias(), alg.scalar_multiplym(learning_rate / n, layer->get_delta())));
		}
	}
}

void MLPPWGAN::update_generator_parameters(Vector<Ref<MLPPMatrix>> hidden_layer_updations, real_t learning_rate) {
	MLPPLinAlg alg;

	if (!network.empty()) {
		for (int i = network.size() / 2; i >= 0; i--) {
			Ref<MLPPHiddenLayer> layer = network[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;
			layer->set_weights(alg.subtractionm(layer->get_weights(), hidden_layer_updations[(network.size() - 2) - i + 1]));
			layer->set_bias(alg.subtract_matrix_rows(layer->get_bias(), alg.scalar_multiplym(learning_rate / n, layer->get_delta())));
		}
	}
}

MLPPWGAN::DiscriminatorGradientResult MLPPWGAN::compute_discriminator_gradients(const Ref<MLPPVector> &y_hat, const Ref<MLPPVector> &output_set) {
	MLPPCost mlpp_cost;
	MLPPActivation avn;
	MLPPLinAlg alg;
	MLPPReg regularization;

	DiscriminatorGradientResult data;

	output_layer->set_delta(alg.hadamard_productnv(mlpp_cost.run_cost_deriv_vector(output_layer->get_cost(), y_hat, output_set), avn.run_activation_deriv_vector(output_layer->get_activation(), output_layer->get_z())));

	data.output_w_grad = alg.mat_vec_multv(alg.transposem(output_layer->get_input()), output_layer->get_delta());
	data.output_w_grad = alg.additionnv(data.output_w_grad, regularization.reg_deriv_termv(output_layer->get_weights(), output_layer->get_lambda(), output_layer->get_alpha(), output_layer->get_reg()));

	if (!network.empty()) {
		Ref<MLPPHiddenLayer> layer = network[network.size() - 1];

		layer->set_delta(alg.hadamard_productm(alg.outer_product(output_layer->get_delta(), output_layer->get_weights()), avn.run_activation_deriv_matrix(layer->get_activation(), layer->get_z())));

		Ref<MLPPMatrix> hidden_layer_w_grad = alg.matmultm(alg.transposem(layer->get_input()), layer->get_delta());

		data.cumulative_hidden_layer_w_grad.push_back(alg.additionm(hidden_layer_w_grad, regularization.reg_deriv_termm(layer->get_weights(), layer->get_lambda(), layer->get_alpha(), layer->get_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:" << layer.weights.size() << "x" << layer.weights[0].size() << std::endl;

		for (int i = network.size() - 2; i > network.size() / 2; i--) {
			layer = network[i];
			Ref<MLPPHiddenLayer> next_layer = network[i + 1];

			layer->set_delta(alg.hadamard_productm(alg.matmultm(next_layer->get_delta(), alg.transposem(next_layer->get_weights())), avn.run_activation_deriv_matrix(layer->get_activation(), layer->get_z())));

			hidden_layer_w_grad = alg.matmultm(alg.transposem(layer->get_input()), layer->get_delta());

			data.cumulative_hidden_layer_w_grad.push_back(alg.additionm(hidden_layer_w_grad, regularization.reg_deriv_termm(layer->get_weights(), layer->get_lambda(), layer->get_alpha(), layer->get_reg()))); // Adding to our cumulative hidden layer grads. Maintain reg terms as well.
		}
	}

	return data;
}

Vector<Ref<MLPPMatrix>> MLPPWGAN::compute_generator_gradients(const Ref<MLPPVector> &y_hat, const Ref<MLPPVector> &output_set) {
	class MLPPCost cost;
	MLPPActivation avn;
	MLPPLinAlg alg;
	MLPPReg regularization;

	Vector<Ref<MLPPMatrix>> cumulative_hidden_layer_w_grad; // Tensor containing ALL hidden grads.

	Ref<MLPPVector> cost_deriv_vector = cost.run_cost_deriv_vector(output_layer->get_cost(), y_hat, output_set);
	Ref<MLPPVector> activation_deriv_vector = avn.run_activation_deriv_vector(output_layer->get_activation(), output_layer->get_z());

	output_layer->set_delta(alg.hadamard_productnv(cost_deriv_vector, activation_deriv_vector));

	Ref<MLPPVector> output_w_grad = alg.mat_vec_multv(alg.transposem(output_layer->get_input()), output_layer->get_delta());
	output_w_grad = alg.additionnv(output_w_grad, regularization.reg_deriv_termv(output_layer->get_weights(), output_layer->get_lambda(), output_layer->get_alpha(), output_layer->get_reg()));

	if (!network.empty()) {
		Ref<MLPPHiddenLayer> layer = network[network.size() - 1];

		Ref<MLPPMatrix> activation_deriv_matrix = avn.run_activation_deriv_matrix(layer->get_activation(), layer->get_z());
		layer->set_delta(alg.hadamard_productm(alg.outer_product(output_layer->get_delta(), output_layer->get_weights()), activation_deriv_matrix));

		Ref<MLPPMatrix> hidden_layer_w_grad = alg.matmultm(alg.transposem(layer->get_input()), layer->get_delta());
		cumulative_hidden_layer_w_grad.push_back(alg.additionm(hidden_layer_w_grad, regularization.reg_deriv_termm(layer->get_weights(), layer->get_lambda(), layer->get_alpha(), layer->get_reg()))); // Adding to our cumulative hidden layer grads. Maintain reg terms as well.

		for (int i = network.size() - 2; i >= 0; i--) {
			layer = network[i];
			Ref<MLPPHiddenLayer> next_layer = network[i + 1];

			activation_deriv_matrix = avn.run_activation_deriv_matrix(layer->get_activation(), layer->get_z());

			layer->set_delta(alg.hadamard_productm(alg.matmultm(next_layer->get_delta(), alg.transposem(next_layer->get_weights())), activation_deriv_matrix));
			hidden_layer_w_grad = alg.matmultm(alg.transposem(layer->get_input()), layer->get_delta());
			cumulative_hidden_layer_w_grad.push_back(alg.additionm(hidden_layer_w_grad, regularization.reg_deriv_termm(layer->get_weights(), layer->get_lambda(), layer->get_alpha(), layer->get_reg()))); // Adding to our cumulative hidden layer grads. Maintain reg terms as well.
		}
	}

	return cumulative_hidden_layer_w_grad;
}

void MLPPWGAN::handle_ui(int epoch, real_t cost_prev, const Ref<MLPPVector> &y_hat, const Ref<MLPPVector> &output_set) {
	MLPPUtilities::cost_info(epoch, cost_prev, cost(y_hat, output_set));

	PLOG_MSG("Layer " + itos(network.size() + 1) + ":");

	MLPPUtilities::print_ui_vb(output_layer->get_weights(), output_layer->get_bias());

	if (!network.empty()) {
		for (int i = network.size() - 1; i >= 0; i--) {
			Ref<MLPPHiddenLayer> layer = network[i];

			PLOG_MSG("Layer " + itos(i + 1) + ":");

			MLPPUtilities::print_ui_mb(layer->get_weights(), layer->get_bias());
		}
	}
}

void MLPPWGAN::_bind_methods() {
	ClassDB::bind_method(D_METHOD("get_output_set"), &MLPPWGAN::get_output_set);
	ClassDB::bind_method(D_METHOD("set_output_set", "val"), &MLPPWGAN::set_output_set);
	ADD_PROPERTY(PropertyInfo(Variant::OBJECT, "output_set", PROPERTY_HINT_RESOURCE_TYPE, "MLPPMatrix"), "set_output_set", "get_output_set");

	ClassDB::bind_method(D_METHOD("get_k"), &MLPPWGAN::get_k);
	ClassDB::bind_method(D_METHOD("set_k", "val"), &MLPPWGAN::set_k);
	ADD_PROPERTY(PropertyInfo(Variant::INT, "k"), "set_k", "get_k");

	ClassDB::bind_method(D_METHOD("generate_example", "n"), &MLPPWGAN::generate_example);
	ClassDB::bind_method(D_METHOD("gradient_descent", "learning_rate", "max_epoch", "ui"), &MLPPWGAN::gradient_descent, false);
	ClassDB::bind_method(D_METHOD("score"), &MLPPWGAN::score);
	ClassDB::bind_method(D_METHOD("save", "file_name"), &MLPPWGAN::save);

	ClassDB::bind_method(D_METHOD("add_layer", "activation", "weight_init", "reg", "lambda", "alpha"), &MLPPWGAN::add_layer, MLPPUtilities::WEIGHT_DISTRIBUTION_TYPE_DEFAULT, MLPPReg::REGULARIZATION_TYPE_NONE, 0.5, 0.5);
	ClassDB::bind_method(D_METHOD("add_output_layer", "weight_init", "reg", "lambda", "alpha"), &MLPPWGAN::add_output_layer, MLPPUtilities::WEIGHT_DISTRIBUTION_TYPE_DEFAULT, MLPPReg::REGULARIZATION_TYPE_NONE, 0.5, 0.5);
}