语义分割混淆矩阵、 mIoU、mPA计算
一、操作
需要会调试代码的人自己改，小白直接运行会出错
这是我从自己的大文件里摘取的一部分代码，可以运行，只是要改的文件地址path比较多，遇到双引号“”的地址注意一下，不然地址不对容易出错
把 calculate.py和 utiles_metrics.py放在同一文件夹下，然后运行 calculate.py。
二、理解
test_mIou,test_mPA,test_miou,test_mpa=compute_mIoU(gt_dir, pred_dir, image_ids, num_classes, name_classes,weight_name) # 执行计算mIoU的函数gt_dir 真实标签文件夹
pred_dir 预测结果文件夹
主要是这两个变量设置，后面的可以选择性修改
image_ids 文件名称 dirList(pred_dir,path_list) saveList(path_list) 这两个函数得到
num_classes 类别数
name_classes 类别名称
weight_name 权重名称
hist为混淆矩阵，mIoU为交并比
三、代码
calculate.py
# -*- coding: utf-8 -*-
import torch
import os
from time import time
# from PIL import Image
from utils_metrics import compute_mIoU
def saveList(pathName):
for file_name in pathName:
#f=open("C:/Users/Administrator/Desktop/DeepGlobe-Road-Extraction-link34-py3/dataset/real/gt.txt", "x")
with open("./dataset/gt.txt", "a") as f:
f.write(file_name.split(".")[0] + "\n")
f.close
def dirList(gt_dir,path_list):
for i in range(0, len(path_list)):
path = os.path.join(gt_dir, path_list[i])
if os.path.isdir(path):
saveList(os.listdir(path))
data_path = './dataset/'
f=open("./dataset/gt.txt", 'w')
gt_dir = os.path.join(data_path, "real/")
pred_dir = "./submits/log01_Dink101_five_100/test_iou/iou_60u/"
path_list = os.listdir(pred_dir)
path_list.sort()
dirList(pred_dir,path_list)
saveList(path_list)
num_classes=2
name_classes = ["nontarget","target"]
weight_name='log01_Dink101_five_100'
image_ids = open(os.path.join(data_path, "gt.txt"),'r').read().splitlines()
test_mIou,test_mPA,test_miou,test_mpa=compute_mIoU(gt_dir, pred_dir, image_ids, num_classes, name_classes,weight_name) # 执行计算mIoU的函数
print(' test_mIoU: '+str(test_miou))
utiles_metrics.py
from os.path import join
import numpy as np
import torch
import torch.nn.functional as F
from PIL import Image
import os
import cv2
# from matplotlib import pyplot as plt
import shutil
import numpy as np
# from matplotlib.pyplot import MultipleLocator
def f_score(inputs, target, beta=1, smooth = 1e-5, threhold = 0.5):
n, c, h, w = inputs.size()
nt, ht, wt, ct = target.size()
if h != ht and w != wt:
inputs = F.interpolate(inputs, size=(ht, wt), mode="bilinear", align_corners=True)
temp_inputs = torch.softmax(inputs.transpose(1, 2).transpose(2, 3).contiguous().view(n, -1, c),-1)
temp_target = target.view(n, -1, ct)
#--------------------------------------------#
# 计算dice系数
#--------------------------------------------#
temp_inputs = torch.gt(temp_inputs, threhold).float()
tp = torch.sum(temp_target[...,:-1] * temp_inputs, axis=[0,1])
fp = torch.sum(temp_inputs , axis=[0,1]) - tp
fn = torch.sum(temp_target[...,:-1] , axis=[0,1]) - tp
score = ((1 + beta ** 2) * tp + smooth) / ((1 + beta ** 2) * tp + beta ** 2 * fn + fp + smooth)
score = torch.mean(score)
return score
# 设标签宽W，长H
def fast_hist(a, b, n):
#--------------------------------------------------------------------------------#
# a是转化成一维数组的标签，形状(H×W,)；b是转化成一维数组的预测结果，形状(H×W,)
#--------------------------------------------------------------------------------#
k = (a >= 0) & (a < n)
#--------------------------------------------------------------------------------#
# np.bincount计算了从0到n**2-1这n**2个数中每个数出现的次数，返回值形状(n, n)
# 返回中，写对角线上的为分类正确的像素点
#--------------------------------------------------------------------------------#
return np.bincount(n * a[k].astype(int) + b[k], minlength=n ** 2).reshape(n, n)
def per_class_iu(hist):
return np.diag(hist) / np.maximum((hist.sum(1) + hist.sum(0) - np.diag(hist)), 1)
def per_class_PA(hist):
return np.diag(hist) / np.maximum(hist.sum(1), 1)
def compute_mIoU(gt_dir, pred_dir, png_name_list, num_classes, name_classes,weight_name):
# print('Num classes', num_classes)
#-----------------------------------------#
# 创建一个全是0的矩阵，是一个混淆矩阵
#-----------------------------------------#
hist = np.zeros((num_classes, num_classes))
#------------------------------------------------#
# 获得验证集标签路径列表，方便直接读取
# 获得验证集图像分割结果路径列表，方便直接读取
#------------------------------------------------#
gt_imgs = [join(gt_dir, x + ".png") for x in png_name_list]
pred_imgs = [join(pred_dir, x + ".png") for x in png_name_list]
# building_iou=[]
# background_iou=[]
m_iou=[]
# building_pa=[]
# background_pa=[]
m_pa=[]
#------------------------------------------------#
# 读取每一个（图片-标签）对
#------------------------------------------------#
for ind in range(len(gt_imgs)):
#------------------------------------------------#
# 读取一张图像分割结果，转化成numpy数组
#------------------------------------------------#
pred = np.array(Image.open(pred_imgs[ind]))
#------------------------------------------------#
# 读取一张对应的标签，转化成numpy数组
#------------------------------------------------#
label = np.array(Image.open(gt_imgs[ind]))
# 如果图像分割结果与标签的大小不一样，这张图片就不计算
if len(label.flatten()) != len(pred.flatten()):
print(
'Skipping: len(gt) = {:d}, len(pred) = {:d}, {:s}, {:s}'.format(
len(label.flatten()), len(pred.flatten()), gt_imgs[ind],
pred_imgs[ind]))
continue
#------------------------------------------------#
# 对一张图片计算21×21的hist矩阵，并累加
#------------------------------------------------#
a=label.flatten()
a//=254
b=pred.flatten()
b//=254
hist += fast_hist(a, b,num_classes)
# # 每计算10张就输出一下目前已计算的图片中所有类别平均的mIoU值
# mIoUs = per_class_iu(hist)
# mPA = per_class_PA(hist)
# m_iou.append(100 * np.nanmean(mIoUs[1]))
# m_pa.append(100 * np.nanmean(mPA[1]))
# # if ind > 0 and ind % 10 == 0:
# # print('{:d} / {:d}: mIou-{:0.2f}; mPA-{:0.2f}'.format(ind, len(gt_imgs),
# # 100 * np.nanmean(mIoUs[1]),
# # 100 * np.nanmean(mPA[1])))
mIoUs = per_class_iu(hist)
mPA = per_class_PA(hist)
print(mIoUs)
# plt.figure()
# x=np.arange(len(m_iou))
# plt.plot(x,m_iou)
# plt.plot(x,m_pa)
# plt.grid(True)
# y_major_locator=MultipleLocator(10)#把y轴的刻度间隔设置为10，并存在变量里
# ax = plt.gca()
# ax.yaxis.set_major_locator(y_major_locator)
# ax.set_ylim(0,100)
# plt.xlabel('Order')
# plt.ylabel('mIOU & mPA')
# plt.legend(['mIOU','mPA'],loc="upper right")
# targ=os.path.join(pred_dir,os.path.pardir)
# plt.savefig(os.path.join(targ, weight_name[:-3]+"_sin_miou.png"))
return m_iou,m_pa,str(round(mIoUs[1] * 100, 2)),str(round(mPA[1] * 100, 2))
调试
