用tensorflow画ROC曲线

来源：互联网发布：电子教室软件哪个好编辑：程序博客网时间：2024/05/21 01:42

1. 先准备好你的数据文件，csv格式，该文件共3列，第一列是数据id，第2列是预测分数（0到1），第3列是数据的label（0或1）

2. 运行下面的python程序：python tf_roc.py /tmp/predict_label.csv 200 /tmp/tb_roc

3. 其中第2个参数200表示画ROC曲线的精度，越大，曲线越精细。

4. 启动tensorboard： tensorboard --logdir=/tmp/tb_roc

5. 打开浏览器http://127.0.0.1:6006/，结果如下图

6. 关于画图方法。其实tensorboard无法画散点图。因此这里采用SummaryWriter的add_summary中第2个参数global_step来替代ROC中的横坐标fpr。globa_step原来是用来表示训练时的迭代次数的。为了能在tb中显示出来，需要把tpr放大100倍，或是1000倍。

7. 代码如下

import sysimport tensorflow as tfclass tf_roc():    def __init__(self, predict_label_file, threshold_num, save_dir):        '''file format: dataid,predict_score,label        the predict_score should be between 0 and 1        the label should be 0 or 1        threshold_num: number of threshold will plot'''        #load predict_label_file to predicts and labels        fd = open(predict_label_file)        fdl = fd.readline()        self.predicts = []        self.labels = []        self.total = 0        while len(fdl) > 0:            val = fdl.split(',')            self.predicts.append(float(val[1]))             self.labels.append(True if int(val[2]) == 1 else False)             fdl = fd.readline()            self.total += 1        fd.close()        self.threshold_num = threshold_num        self.trues = 0 #total of True labels         self.fpr = [] #false positive        self.tpr = [] #true positive        self.ths = [] #thresholds        self.save_dir = save_dir        self.writer = tf.train.SummaryWriter(self.save_dir)    def calc(self):        for label in self.labels:            if label:                self.trues += 1        threshold_step = 1. / self.threshold_num        for t in range(self.threshold_num + 1):            th = 1 - threshold_step * t            tn, tp, fp, fpr, tpr = self._calc_once(th)            self.fpr.append(fpr)            self.tpr.append(tpr)            self.ths.append(th)            self._save(fpr, tpr)        print self.fpr        print self.tpr        print self.ths    def _save(self, fpr, tpr):        summt = tf.Summary()        summt.value.add(tag="roc", simple_value = tpr)        self.writer.add_summary(summt, fpr * 100) #for tensorboard step drawable        self.writer.flush()    def _calc_once(self, t):        fp = 0        tp = 0        tn = 0        for i in range(self.total):            if not self.labels[i]:                if self.predicts[i] >= t:                    fp += 1                else:                    tn += 1            elif self.predicts[i] >= t:                tp += 1        #fpr = fp / float(fp + tn) #precision        fpr = fp / float(fp + tp) #detection        tpr = tp / float(self.trues)        return tn, tp, fp, fpr, tpr        if __name__ == '__main__':    predict_label_file, threshold_num, save_dir = sys.argv[1:4]    roc = tf_roc(predict_label_file, int(threshold_num), save_dir)    roc.calc()

1 0