SupportVectorRegression.py 文件源码

def train(self):

        if (self.status != 'init'):
            print("Please load train data and init W first.")
            return self.W

        self.status = 'train'

        original_X = self.train_X[:, 1:]

        K = utility.Kernel.kernel_matrix(self, original_X)

        # P = Q, q = p, G = -A, h = -c

        P = cvxopt.matrix(np.bmat([[K, -K], [-K, K]]))
        q = cvxopt.matrix(np.bmat([self.epsilon - self.train_Y, self.epsilon + self.train_Y]).reshape((-1, 1)))
        G = cvxopt.matrix(np.bmat([[-np.eye(2 * self.data_num)], [np.eye(2 * self.data_num)]]))
        h = cvxopt.matrix(np.bmat([[np.zeros((2 * self.data_num, 1))], [self.C * np.ones((2 * self.data_num, 1))]]))
        # A = cvxopt.matrix(np.append(np.ones(self.data_num), -1 * np.ones(self.data_num)), (1, 2*self.data_num))
        # b = cvxopt.matrix(0.0)
        cvxopt.solvers.options['show_progress'] = False
        solution = cvxopt.solvers.qp(P, q, G, h)

        # Lagrange multipliers
        alpha = np.array(solution['x']).reshape((2, -1))
        self.alpha_upper = alpha[0]
        self.alpha_lower = alpha[1]
        self.beta = self.alpha_upper - self.alpha_lower

        sv = abs(self.beta) > 1e-5
        self.sv_index = np.arange(len(self.beta))[sv]
        self.sv_beta = self.beta[sv]
        self.sv_X = original_X[sv]
        self.sv_Y = self.train_Y[sv]

        free_sv_upper = np.logical_and(self.alpha_upper > 1e-5, self.alpha_upper < self.C)
        self.free_sv_index_upper = np.arange(len(self.alpha_upper))[free_sv_upper]
        self.free_sv_alpha_upper = self.alpha_upper[free_sv_upper]
        self.free_sv_X_upper = original_X[free_sv_upper]
        self.free_sv_Y_upper = self.train_Y[free_sv_upper]

        free_sv_lower = np.logical_and(self.alpha_lower > 1e-5, self.alpha_lower < self.C)
        self.free_sv_index_lower = np.arange(len(self.alpha_lower))[free_sv_lower]
        self.free_sv_alpha_lower = self.alpha_lower[free_sv_lower]
        self.free_sv_X_lower = original_X[free_sv_lower]
        self.free_sv_Y_lower = self.train_Y[free_sv_lower]

        short_b_upper = self.free_sv_Y_upper[0] - np.sum(self.sv_beta * utility.Kernel.kernel_matrix_xX(self, self.free_sv_X_upper[0], self.sv_X)) - self.epsilon
        short_b_lower = self.free_sv_Y_lower[0] - np.sum(self.sv_beta * utility.Kernel.kernel_matrix_xX(self, self.free_sv_X_lower[0], self.sv_X)) + self.epsilon

        self.sv_avg_b = (short_b_upper + short_b_lower) / 2

        return self.W