convnet.py 文件源码

def find_optimal_C_for_AUC(xTrain, yTrain, xTest, yTest):
    C_2d_range = [10.0 ** i for i in range(-3, 3)]

    accuracy = np.array([])
    auc_score = np.array([])

    for Ctry in C_2d_range:
        clf = SVC(C=Ctry, kernel="linear", probability=True)
        clf.fit(xTrain, yTrain)
        pred = clf.predict(xTest)
        pred_proba = clf.predict_proba(xTest)
        accuracy = np.append(accuracy, np.average(yTest == pred))
        auc_score = np.append(auc_score,
                              roc_auc_score(yTest, pred_proba[:, 1]))
        print "C: {}" .format(Ctry)
        print "accuracy: {}" .format(accuracy[-1])
        print "AUC: {}" .format(auc_score[-1])

    # Extract the optimal parameters to train the final model
    best_auc_idx = np.where(auc_score == max(auc_score))[0]
    best_acc_idx = np.where(accuracy == max(accuracy[best_auc_idx]))[0]
    best_C = C_2d_range[best_acc_idx[0]]

    return best_C