Python绘制ROC曲线

ROC曲线全程为受试者工作特性（Receiver Operating Characteristic)曲线，可以用来评价学习器的泛化性能。首先我们根据学习器的预测结果对样例进行排序，按此顺序逐个把样本作为正例进行预测，每次计算出两个重要的值，分别以他们为横纵坐标作图，就得到了ROC曲线。ROC曲线的横轴为假正例率(FPR), 纵轴为真正例率(TPR),计算公式分别为：

FPR = FP/(FP+TN);            TPR=TP/(TP+FN).

程序分为三个部分：1.获得学习器的输出值    2.根据输出求出FPR和TPR    3.绘制ROC图

一 、获得学习器的输出

x = tf.placeholder(dtype=tf.float32, shape=[1, 224, 224, 3])
sess = tf.InteractiveSession()

#build the network
arg_scope = alexnet.alexnet_v2_arg_scope()
with slim.arg_scope(arg_scope):
    logits, _ = alexnet.alexnet_v2(x, num_classes=2, is_training=False)
'''
model = net.Alexnet("pre_trained/alexnet.npy")
logits = model.build(x)
'''
saver = tf.train.Saver()
tf.global_variables_initializer().run()

#restore the pre-trained weights
saver.restore(sess, "logs/alexnet_2000.ckpt")

def get_test_result(sess, path, true_result=0, begin=0, end=1000):
    error = 0
    lit = []

    for parent,_,filenames in os.walk(path):
        filenames.sort(key=lambda x: int(x[:-4]))
        filenames = filenames[begin:end]

        for filename in filenames:
            file_path = parent + "/" + filename

            #preprocess the images, by inception v1, image=2*(image/255.0) - 1.0, scale the image to (-1,1)
            image = cv2.resize(cv2.imread(file_path), (224,224))
            image = 2*(image / 255.0) - 1.0
            img = np.reshape(image, [-1,224,224,3])

            # calculate the logits
            predict = sess.run(logits,feed_dict={x:img})
            predict = np.reshape(predict, [-1])

            result = np.argmax(predict, axis=0)

            if result!=true_result:
                print(predict, file_path)
                error += 1
            else:
                print(predict)

            lit.append(predict[0])

        # sort the list from small to large, if want to get the reverse result, use lit.sort(reverse=True)
        lit.sort()
        err = error/len(filenames)

        return lit ,err

#get the normal and polyp score list
normal_list, err_normal = get_test_result(sess, "/home/hdl/ALL_IMAGE/dataset/normal", true_result=1)
polyp_list, err_polyp = get_test_result(sess, "/home/hdl/ALL_IMAGE/dataset/polyp", true_result=0, begin=5000, end=-1)

#reverse the list
normal_list = normal_list[::-1]
polyp_list = polyp_list[::-1]

#get the all score list used to preduct threadhold
all_list = normal_list + polyp_list
all_list.sort(reverse=True)

二、根据输出求出FPR和TPR

normal = np.array(normal_list)
polyp = np.array(polyp_list)

TPR = []
FPR = []
for threadhold in all_list:
    temp_polyp = polyp >= threadhold
    temp_normal = normal >= threadhold

    tp = np.sum(temp_polyp == 1)
    fn = np.sum(temp_polyp == 0)

    tn = np.sum(temp_normal == 0)
    fp = np.sum(temp_normal == 1)

    #the code above can be replaced by this
    '''
    tp = np.sum(temp_polyp >= threadhold)
    '''

    tpr = tp/(tp+fn)
    fpr = fp/(tn+fp)

    TPR.append(tpr)
    FPR.append(fpr)

#combine the data and saved as .npy
arr = np.concatenate((FPR,TPR), axis=0)
np.save("Alexnet_scratch.npy", arr)

三、绘制ROC曲线

import numpy as np
from matplotlib import pyplot as plot

alexnet_finetune = np.load("Alexnet_finetune.npy")
Alexnet_scratch = np.load("Alexnet_scratch.npy")
VGG = np.load("VGG.npy")

alexnet_finetune = np.reshape(alexnet_finetune, newshape=[2,-1])
Alexnet_scratch = np.reshape(Alexnet_scratch, newshape=[2,-1])
VGG = np.reshape(VGG, newshape=[2,-1])

plot.title("Receive operating characteristic curve")
plot.xlabel("False Positive Rate")
plot.ylabel("True Positive Rate")

plot.plot(alexnet_finetune[0],alexnet_finetune[1], color="red", label="Alexnet finetune")
plot.plot(Alexnet_scratch[0], Alexnet_scratch[1], color="blue", label="Alexnet scratch")
plot.plot(VGG[0], VGG[1], color="green", label="VGG with global avgpool")

plot.legend()
plot.show()