Monitor.py 文件源码

def optimize_training_parameters(self, n):
        # data
        from_timestamp = self.min_timestamp
        to_timestamp = self.min_timestamp + datetime.timedelta(days=365) + datetime.timedelta(hours=1)
        train_timestamps, train_values = self.load_monitor_data(from_timestamp, to_timestamp, "1")
        train_data = np.array(train_values)[:, 0:5]

        # parameters
        nu = np.linspace(start=1e-5, stop=1e-2, num=n)
        gamma = np.linspace(start=1e-6, stop=1e-3, num=n)
        opt_diff = 1.0
        opt_nu = None
        opt_gamma = None
        fw = open("training_param.csv", "w")
        fw.write("nu,gamma,diff\n")
        for i in range(len(nu)):
            for j in range(len(gamma)):
                classifier = svm.OneClassSVM(kernel="rbf", nu=nu[i], gamma=gamma[j])
                classifier.fit(train_data)
                label = classifier.predict(train_data)
                p = 1 - float(sum(label == 1.0)) / len(label)
                diff = math.fabs(p-nu[i])
                if diff < opt_diff:
                    opt_diff = diff
                    opt_nu = nu[i]
                    opt_gamma = gamma[j]
                fw.write(",".join([str(nu[i]), str(gamma[j]), str(diff)]) + "\n")
        fw.close()
        return opt_nu, opt_gamma