本文主要是在WIDERFACE数据集上评估自己的模型（利用官网代码生成的模型），包括图片的检测、损失函数的绘制、验证集上准确率和召回率的计算以及P-R曲线的绘制。
单图片检测部分是比较简单的一个步骤，忘记了源代码中是否有保存检测结果的代码，如果没有自己添加即可。
./darknet detect cfg/yolov3.cfg yolov3.weights data/dog.jpg
这部分的前提应该是你有每次迭代过程的记录文件，可以是nohup命令得出的nohup.out文件，也可以是其他形式的文件。
# coding=utf-8
# 该文件用来提取训练log，去除不可解析的log后使log文件格式化，生成新的log文件供可视化工具绘图
import inspect
import os
import random
import sys
def extract_log(log_file,new_log_file,key_word):
with open(log_file, 'r') as f:
with open(new_log_file, 'w') as train_log:
#f = open(log_file)
#train_log = open(new_log_file, 'w')
for line in f:
# 去除多gpu的同步log
if 'Syncing' in line:
continue
# 去除除零错误的log
if 'nan' in line:
continue
if key_word in line:
train_log.write(line)
f.close()
train_log.close()
extract_log('train_yolov3.log','train_log_loss.txt','images')
extract_log('train_yolov3.log','train_log_iou.txt','IOU')
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
#%matplotlib inline
lines =50000 #改为自己生成的train_log_loss.txt中的行数
result = pd.read_csv('train_log_loss.txt', skiprows=[x for x in range(lines) if ((x%10!=9) |(x<1000))] ,error_bad_lines=False, names=['loss', 'avg', 'rate', 'seconds', 'images'])
result.head()
result['loss']=result['loss'].str.split(' ').str.get(1)
result['avg']=result['avg'].str.split(' ').str.get(1)
result['rate']=result['rate'].str.split(' ').str.get(1)
result['seconds']=result['seconds'].str.split(' ').str.get(1)
result['images']=result['images'].str.split(' ').str.get(1)
result.head()
result.tail()
# print(result.head())
# print(result.tail())
# print(result.dtypes)
print(result['loss'])
print(result['avg'])
print(result['rate'])
print(result['seconds'])
print(result['images'])
result['loss']=pd.to_numeric(result['loss'])
result['avg']=pd.to_numeric(result['avg'])
result['rate']=pd.to_numeric(result['rate'])
result['seconds']=pd.to_numeric(result['seconds'])
result['images']=pd.to_numeric(result['images'])
result.dtypes
fig = plt.figure()
ax = fig.add_subplot(1, 1, 1)
ax.plot(result['avg'].values,label='avg_loss')
# ax.plot(result['loss'].values,label='loss')
ax.legend(loc='best') #图列自适应位置
ax.set_title('The loss curves')
ax.set_xlabel('batches')
fig.savefig('avg_loss')
# fig.savefig('loss')
这是我第一个使用的计算准确率的方法，虽然处理预测文件的时候，我去除了非人脸信息的检测结果，也设置了相应的阈值，但是很奇怪的是计算出的准确率离奇的低。
从上述参考资源处下载voc_eval.py和reval_voc_py3.py文件
利用官网代码对验证集的图片进行检测，默认生成comp4_det_test_face.txt(在results文件夹下)
./darknet detector valid cfg/voc.data cfg/yolov3-voc.cfg backup/yolov3-voc_final.weights -out car.txt -gpu 0 -thresh .5
每次检测后都会在voc2007数据集文件夹下生成一个annotation的文件，需要再次检测时，得把它删除，不然报错。
python3.6 reval_voc_py3.py --voc_dir /home/amax/cxw/darknet-master/voc/VOCdevkit --year 2007 --image_set test --classes data/voc-my.names Testforpr
上一步骤中，已经求得了准确率和召回率的值，并保存在了pkl文件中，绘制p-r曲线时，只需要加载上这个文件即可。
# coding=utf-8
import _pickle as cPickle
import matplotlib.pyplot as plt
plt.rcParams['font.sans-serif'] = [u'NSimSun']
plt.rcParams['axes.unicode_minus'] = False
fr = open('tower_pr.pkl', 'rb') # 这里open中第一个参数需要修改成自己生产的pkl文件
inf = cPickle.load(fr)
fr.close()
# fbad = open('tower_pr.pkl','rb')#这里open中第一个参数需要修改成自己生产的pkl文件
# inf_bad = cPickle.load(fbad)
# fbad.close()
x = inf['rec']
y = inf['prec']
plt.figure()
plt.xlabel('召回率', size=15)
plt.ylabel('精确率', size=15)
plt.xticks(fontproperties='Times New Roman', size=14)
plt.yticks(fontproperties='Times New Roman', size=14)
plt.plot(x, y)
plt.savefig("PR曲线.svg", bbox_inches='tight') # plt保存需要在show之前
plt.show()
print('AP：', inf['ap'])
AP值的计算无非就是将验证集的检测结果同原数据进行对比，设置相应的阈值，计算数据即可。
import os
# txt_file为配置文件.data中的valid
txt_file = '/home/amax/cxw/darknet-master/scripts/2007_val.txt'
f = open(txt_file)
lines = f.readlines()
for line in lines:
line = line.split('/')[-1][0:-5]
# test_out_file 为转换后保存的结果地址
test_out_file = '/home/amax/cxw/darknet-master/testresults'
# 下面3个with需要自己的修改，修改成自己对应的类别
with open(os.path.join(test_out_file , line + '.txt'), "a") as new_f:
f1 = open('/home/amax/cxw/darknet-master/results/comp4_det_test_face.txt', 'r')
f1_lines = f1.readlines()
for f1_line in f1_lines:
f1_line = f1_line.split()
if line == f1_line[0]:
new_f.write("%s %s %s %s %s %s\n" % ('smoke', f1_line[1], f1_line[2], f1_line[3], f1_line[4], f1_line[5]))
import os
from PIL import Image
import numpy as np
# label_img为数据集的labels地址，img_path为数据集images的地址
label_img = '/home/amax/cxw/darknet-master/voc/VOCdevkit/VOC2007/labels'
img_path = '/home/amax/cxw/darknet-master/voc/VOCdevkit/VOC2007/JPEGImages'
classes = {
0:'smoke',
1:'white',
2:'red'
}
for line in lines:
line = line.split('/')[-1][0:-5] + '.txt'
txt = label_img + line
img = np.array(Image.open(img_path + line.split('/')[-1][0:-4] + '.jpg'))
sh, sw = img.shape[0], img.shape[1]
# gt_out_file为转换后的地址
gt_out_file = '/home/amax/cxw/darknet-master/yoloresults'
with open(os.path.join(gt_out_file , line ), "a") as new_f:
f1 = open(txt)
f1_lines = f1.readlines()
for f1_line in f1_lines:
f1_line = f1_line.split()
x = float(f1_line[1]) * sw
y = float(f1_line[2]) * sh
w = float(f1_line[3]) * sw
h = float(f1_line[4]) * sh
xmin = x+1-w/2
ymin = y+1-h/2
xmax = x+1+w/2
ymax = y+1+h/2
new_f.write("%s %s %s %s %s\n" % (classes[int(f1_line[0])], xmin ,ymin,xmax,ymax))
import os
def compute_IOU(rec1, rec2):
"""
计算两个矩形框的交并比。
:param rec1: (x0,y0,x1,y1) (x0,y0)代表矩形左上的顶点，（x1,y1）代表矩形右下的顶点。下同。
:param rec2: (x0,y0,x1,y1)
:return: 交并比IOU.
"""
left_column_max = max(rec1[0], rec2[0])
right_column_min = min(rec1[2], rec2[2])
up_row_max = max(rec1[1], rec2[1])
down_row_min = min(rec1[3], rec2[3])
# 两矩形无相交区域的情况
if left_column_max >= right_column_min or down_row_min <= up_row_max:
return 0
# 两矩形有相交区域的情况
else:
S1 = (rec1[2] - rec1[0]) * (rec1[3] - rec1[1])
S2 = (rec2[2] - rec2[0]) * (rec2[3] - rec2[1])
S_cross = (down_row_min - up_row_max) * (right_column_min - left_column_max)
return S_cross / (S1 + S2 - S_cross)
# gt为yolo数据转换后的地址
gt = '/home/amax/cxw/darknet-master/yoloresults/'
# test为检测结果转换后的地址
test = '/home/amax/cxw/darknet-master/testresults/'
# count_gt为标注的所有数据框
count_gt = {
}
# count_test为检测的所有数据框
count_test = {
}
# count_yes_test为检测正确的数据框
count_yes_test = {
}
# count_no_test为检测错误的数据框
count_no_test = {
}
# 计数
for gt_ in os.listdir(gt):
txt = gt + gt_
f = open(txt)
lines = f.readlines()
for line in lines:
line = line.split()
name = line[0]
if name not in count_gt:
count_gt[name] = 0
count_gt[name] += 1
for test_ in os.listdir(test):
txt = test + test_
f = open(txt)
lines = f.readlines()
for line in lines:
line = line.split()
name = line[0]
if name not in count_test:
count_test[name] = 0
count_test[name] += 1
# 下面主要思想：遍历test结果，再遍历对应gt的结果，如果两个框的iou大于一定的阙址并且类别相同，视为正确
for test_ in os.listdir(test):
f_test_txt = test + test_
f_test = open(f_test_txt)
f_test_lines = f_test.readlines()
for f_test_line in f_test_lines:
f_test_line = f_test_line.split()
f_gt_txt = gt + test_
f_gt = open(f_gt_txt)
f_gt_lines = f_gt.readlines()
flag = 1
for f_gt_line in f_gt_lines:
f_gt_line = f_gt_line.split()
IOU = compute_IOU([float(f_gt_line[1]), float(f_gt_line[2]), float(f_gt_line[3]), float(f_gt_line[4])],
[float(f_test_line[2]), float(f_test_line[3]), float(f_test_line[4]), float(f_test_line[5])])
if f_gt_line[0] == f_test_line[0] and IOU >= 0.5 and float(f_test_line[1]) >= 0.3:
flag = 0
if f_test_line[0] not in count_yes_test:
count_yes_test[f_test_line[0]] = 0
count_yes_test[f_test_line[0]] += 1
if flag == 1:
if f_test_line[0] not in count_no_test:
count_no_test[f_test_line[0]] = 0
count_no_test[f_test_line[0]] += 1
# 有以下4个结果，就可以计算相关指标了
print(count_gt)
print(count_test)
print(count_yes_test)
print(count_no_test)