Added Stan Math Support to Shogun (#4216)

cmake/external/StanMath.cmake

@@ -0,0 +1,22 @@
+set(STAN_STABLE_RELEASE 4b1a10bc877d941bbe0a12c198526807be27167a)
+
+set(STAN_INCLUDE_DIR_STAN_MATH ${CMAKE_BINARY_DIR}/StanMath/src/StanMath)
+set(STAN_INCLUDE_DIR_BOOST ${CMAKE_BINARY_DIR}/StanMath/src/StanMath/lib/boost_1.64.0)
+set(STAN_INCLUDE_DIR_CVODES ${CMAKE_BINARY_DIR}/StanMath/src/StanMath/lib/cvodes_2.9.0/include)
+
+include(ExternalProject)
+ExternalProject_Add(
+	StanMath
+	PREFIX ${CMAKE_BINARY_DIR}/StanMath
+	DOWNLOAD_DIR ${THIRD_PARTY_DIR}/StanMath
+	GIT_REPOSITORY https://github.com/stan-dev/math
+	GIT_TAG ${STAN_STABLE_RELEASE}
+	CONFIGURE_COMMAND ""
+	BUILD_COMMAND ""
+	INSTALL_COMMAND
+		${CMAKE_COMMAND} -E copy_if_different ${STAN_INCLUDE_DIR_STAN_MATH} include/shogun/third_party/Stan
+		&& ${CMAKE_COMMAND} -E copy_if_different ${STAN_INCLUDE_DIR_BOOST} include/shogun/third_party/Stan_Boost
+		&& ${CMAKE_COMMAND} -E copy_if_different ${STAN_INCLUDE_DIR_CVODES} include/shogun/third_party/Stan_Cvodes
+	LOG_DOWNLOAD ON
+	)
+add_dependencies(libshogun StanMath)

src/shogun/CMakeLists.txt

@@ -217,6 +217,22 @@ CHECK_FUNCTION_EXISTS(lgammal HAVE_LGAMMAL)
 
 ############################ external dependencies
 
+#Stan math
+include(external/StanMath)
+
+SHOGUN_INCLUDE_DIRS(SCOPE PUBLIC SYSTEM
+  $<BUILD_INTERFACE:${STAN_INCLUDE_DIR_STAN_MATH}>
+  $<INSTALL_INTERFACE:include/shogun/third_party/Stan>
+)
+SHOGUN_INCLUDE_DIRS(SCOPE PUBLIC SYSTEM
+  $<BUILD_INTERFACE:${STAN_INCLUDE_DIR_BOOST}>
+  $<INSTALL_INTERFACE:include/shogun/third_party/Stan_Boost>
+)
+SHOGUN_INCLUDE_DIRS(SCOPE PUBLIC SYSTEM
+  $<BUILD_INTERFACE:${STAN_INCLUDE_DIR_CVODES}>
+  $<INSTALL_INTERFACE:include/shogun/third_party/Stan_Cvodes>
+)
+
 ####### LINALG
 FIND_PACKAGE(Eigen3 ${EIGEN_VERSION_MINIMUM})
 IF(NOT EIGEN3_FOUND)

tests/unit/StanMath/StanPerceptronTest_unittest.cc

@@ -0,0 +1,73 @@
+// tests/unit/stanmath/StanPerceptronTest_unittest.cc
+#include <Eigen/Dense>
+#include <cmath>
+#include <gtest/gtest.h>
+#include <iostream>
+#include <random>
+#include <stan/math.hpp>
+using namespace std;
+using namespace Eigen;
+using namespace stan::math;
+
+/*
+This test will implement a perceptron using stan math
+The perceptron will have one input, the column vector [1 1 1]^T
+Then, it will have 2x3 Matrix of weights that it will learn
+Then, the output is a 2x1 column vector
+In this example, we want to learn the weights W such that the square
+Error loss from the output of the perceptron to [1 1]^T is minimized.
+Since we can find weights from [1 1 1]^T to [1 1]^T in a perceptron,
+this error should be very close to zero after 100 epochs.
+*/
+TEST(StanPerceptronTest, sample_perceptron)
+{
+	// Initialize the Input Vector
+	Matrix<var, 3, 1> inp;
+	inp(0, 0) = 1;
+	inp(1, 0) = 1;
+	inp(2, 0) = 1;
+
+	// Randomly Initialize the weights on the perceptron
+	std::random_device rd{};
+	std::mt19937 gen{rd()};
+	normal_distribution<> d{0, 1};
+	Matrix<var, 2, 3> W1;
+	for (int i = 0; i < 2; ++i)
+	{
+		for (int j = 0; j < 3; ++j)
+		{
+			W1(i, j) = 0.01 * d(gen);
+		}
+	}
+
+	// Define the outputs of the neural network
+	Matrix<var, 2, 1> outputs;
+
+	double learning_rate = 0.1;
+	double last_error = 0;
+	for (int epoch = 0; epoch < 100; ++epoch)
+	{
+		var error = 0;
+		outputs = W1 * inp;
+		for (int i = 0; i < 2; ++i)
+		{
+			error += (outputs(i, 0) - 1) * (outputs(i, 0) - 1);
+		}
+		error.grad();
+
+		// Now use gradient descent to change the weights
+		for (int i = 0; i < 2; ++i)
+		{
+			for (int j = 0; j < 3; ++j)
+			{
+				W1(i, j) = W1(i, j) - learning_rate * W1(i, j).adj();
+			}
+		}
+
+		// Store the value of current error in last_error
+		last_error = value_of(error);
+	}
+
+	// Error should be very close to 0.0
+	EXPECT_NEAR(last_error, 0.0, 1e-6);
+}