如何在Python中解决此问题(为Infogain Loss层创建权重)? [英] How to solve this issue in Python (creating weights for Infogain Loss layer)?

问题描述

我正在使用CNN进行语义分割.我每个班级都有不平衡的像素数. 基于此链接，我我正在尝试创建权重矩阵H，以便为我的失衡类成员定义Infogain loss layer. 我的数据有五个类.我在python中编写了以下代码: 读取示例图像:

I am working on semantic segmentation using CNNs. I have an imbalance number of pixels for each class. Based on this link, I am trying to create weight matrix H in order to define Infogain loss layer for my imbalance class members. My data has five classes. I wrote the following code in python: Reads a sample image:

im=imread(sample_img_path)

计算每个类别的像素数

cl0=np.count_nonzero(im == 0)   #0=background class
.
.
cl4=np.count_nonzero(im == 4)   #4=class 4

输出: 39817 13751 1091 10460 417

output: 39817 13751 1091 10460 417

 #Inverse class weights
    #FORMULA=(total number of sample)/((number of classes)*(number of sample in class i))
    w0=round(sum_/(no_classes*cl0),3)
    w1=round(sum_/(no_classes*cl1),3)
    w2=round(sum_/(no_classes*cl2),3)
    w3=round(sum_/(no_classes*cl3),3)
    w4=round(sum_/(no_classes*cl4),3)
    print w0,w1,w2,w3,w4
L_1=[w0,w1,w2,w3,w4]
    #weighting based on the number of pixel
print L_1
L=[round(i/sum(L_1),2) for i in L_1]  #normalizing the weights
print L
print sum(L)
#creating the H matrix
H=np.eye(5)
print H
#H = np.eye( L, dtype = 'f4' ) 
d=np.diag_indices_from(H)
H[d]=L


print H

blob = caffe.io.array_to_blobproto(H.reshape((1,1,L,L)))
with open( 'infogainH.binaryproto', 'wb' ) as f :
    f.write( blob.SerializeToString() )

print f

除去一些不重要的行后，输出如下:

The output, after removing some unimportant lines, is as follows:

(256, 256)
39817 13751 1091 10460 417
0.329 0.953 12.014 1.253 31.432
<type 'list'>
[0.329, 0.953, 12.014, 1.253, 31.432]
[0.01, 0.02, 0.26, 0.03, 0.68]
1.0
[[ 1.  0.  0.  0.  0.]
 [ 0.  1.  0.  0.  0.]
 [ 0.  0.  1.  0.  0.]
 [ 0.  0.  0.  1.  0.]
 [ 0.  0.  0.  0.  1.]]
[[ 0.01  0.    0.    0.    0.  ]
 [ 0.    0.02  0.    0.    0.  ]
 [ 0.    0.    0.26  0.    0.  ]
 [ 0.    0.    0.    0.03  0.  ]
 [ 0.    0.    0.    0.    0.68]]
Traceback (most recent call last):
  File "create_class_prob.py", line 59, in <module>
    blob = caffe.io.array_to_blobproto(H.reshape((1,1,L,L)))
TypeError: an integer is required

可以看出，它给出了一个错误.我的问题可以分为两部分:

As it can be seen, it is giving an error. My question can be folded into two parts:

1. 如何解决此错误? 我将L替换为5，如下所示:

1. How to solve this error? I replaced L with 5 as follows:

blob = caffe.io.array_to_blobproto(H.reshape((1,1,5,5)))

现在，它没有给出错误，最后一行显示了这一点:

Now, it is not giving error and last line shows this:

<closed file 'infogainH.binaryproto', mode 'wb' at 0x7f94b5775b70>

它创建了文件infogainH.binaryproto，对吗?

1. 对于数据库中的所有图像，此矩阵H是否应为常数?

我非常感谢您的帮助.

I really appreciate any help.

谢谢

推荐答案

1. 您有一个简单的复制粘贴"错误.您从此答案复制了代码，其中L是代表类数的整数.另一方面，在您的代码中，L是具有类权重的 list .在代码中用5替换L确实可以解决问题.

1. You have a simple "copy-paste" bug. You copied your code from this answer where L was an integer representing the number of classes. In your code, on the other hand, L is a list with the class weights. replacing L with 5 in your code does indeed solves the problem.

H应该是常数吗?确实由您决定.

Should H be constant? This is really up to you to decide.

顺便说一句，AFAIK，当前的Caffe版本不支持逐像素信息丢失，您可能需要使用 PR#3855 .

BTW, AFAIK, current caffe version does not support pixel-wise infogain loss, you might need to use the code in PR #3855.

这篇关于如何在Python中解决此问题(为Infogain Loss层创建权重)?的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！