Keras 分割网络自定义评估函数 - mean iou

本文最后更新于:2022年7月4日 上午

Keras训练网络过程中需要实时观察性能,mean iou不是keras自带的评估函数,tf的又觉得不好用,自己写了一个,经过测试没有问题,本文记录自定义keras mean iou评估的实现方法。

计算 IoU

用numpy计算的,作为IoU的ground truth用作测试使用:

1
2
3
4
5
6
def iou_numpy(y_true,y_pred):

intersection = np.sum(np.multiply(y_true.astype('bool'),y_pred == 1))
union = np.sum((y_true.astype('bool')+y_pred.astype('bool'))>0)

return intersection/union

keras metric IoU

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
def iou_keras(y_true, y_pred):
"""
Return the Intersection over Union (IoU).
Args:
y_true: the expected y values as a one-hot
y_pred: the predicted y values as a one-hot or softmax output
Returns:
the IoU for the given label
"""
label = 1
# extract the label values using the argmax operator then
# calculate equality of the predictions and truths to the label
y_true = K.cast(K.equal(y_true, label), K.floatx())
y_pred = K.cast(K.equal(y_pred, label), K.floatx())
# calculate the |intersection| (AND) of the labels
intersection = K.sum(y_true * y_pred)
# calculate the |union| (OR) of the labels
union = K.sum(y_true) + K.sum(y_pred) - intersection
# avoid divide by zero - if the union is zero, return 1
# otherwise, return the intersection over union
return K.switch(K.equal(union, 0), 1.0, intersection / union)

计算 mean IoU

mean IoU 简便起见,选取 (0,1,0.05) 作为不同的IoU阈值,计算平均IoU

numpy 真实值计算

1
2
3
4
5
6
7
8
9
def mean_iou_numpy(y_true,y_pred):

iou_list = []
for thre in list(np.arange(0.0000001,0.99,0.05)):
y_pred_temp = y_pred >= thre
iou = iou_numpy(y_true, y_pred_temp)
iou_list.append(iou)

return np.mean(iou_list)

Keras mean IoU

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
def mean_iou_keras(y_true, y_pred):
"""
Return the mean Intersection over Union (IoU).
Args:
y_true: the expected y values as a one-hot
y_pred: the predicted y values as a one-hot or softmax output
Returns:
the mean IoU
"""
label = 1
# extract the label values using the argmax operator then
# calculate equality of the predictions and truths to the label
y_true = K.cast(K.equal(y_true, label), K.floatx())

mean_iou = K.variable(0)

thre_list = list(np.arange(0.0000001,0.99,0.05))

for thre in thre_list:

y_pred_temp = K.cast(y_pred >= thre, K.floatx())
y_pred_temp = K.cast(K.equal(y_pred_temp, label), K.floatx())
# calculate the |intersection| (AND) of the labels
intersection = K.sum(y_true * y_pred_temp)
# calculate the |union| (OR) of the labels
union = K.sum(y_true) + K.sum(y_pred_temp) - intersection
iou = K.switch(K.equal(union, 0), 1.0, intersection / union)
mean_iou = mean_iou + iou

return mean_iou / len(thre_list)

测试

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
## 随机生成预测值
y_true_np = np.ones([10,10])
y_pred_np = np.random.rand(10,10)

## 真实IoU值
print(f' iou : {iou_numpy(y_true_np, y_pred_np)}')
print(f' mean_iou_numpy : {mean_iou_numpy(y_true_np, y_pred_np)}')

y_true = tf.Variable(y_true_np)
y_pred = tf.Variable(y_pred_np)

## 计算节点
iou_res = iou_keras (y_true, y_pred)
m_iou_res = mean_iou_keras (y_true, y_pred)

## 变量初始化
init_op = tf.global_variables_initializer()

with tf.Session() as sess:
sess.run(init_op)

## 由于存在误差,结果在0.0000001范围内即可认为相同

result = sess.run(iou_res)
print(f'result : {result} \nsame with ground truth: {abs(iou_numpy(y_true_np, y_pred_np) - result)< 0.0000001}')

result = sess.run(m_iou_res)
print(f'result : {result} \nsame with ground truth: {abs(mean_iou_numpy(y_true_np, y_pred_np) - result) < 0.0000001}')

输出:

1
2
3
4
5
6
iou : 0.0
mean_iou_numpy : 0.5295
result : 0.0
same with ground truth: True
result : 0.5295000076293945
same with ground truth: True

源码下载

https://github.com/zywvvd/Python_Practise


Keras 分割网络自定义评估函数 - mean iou
https://www.zywvvd.com/notes/study/deep-learning/keras/my-mean-iou/my-mean-iou/
作者
Yiwei Zhang
发布于
2020年5月25日
许可协议