如何计算map？
1.VOCdevkit/VOC2007/ImageSets/Main/test.txt、val.txt、train.txt、trainval.txt由voc2yolo4.py生成
可以使用trainval_percent和train_percent的比例改变来修改
实验比例的话是trainval_percent=0.9,train_percent=1
根据我来看的代码，使用k-means聚类之后得到的结果不需要改变这边。
————————————2021.4.18——————————————————
D:\ProgramData\Anaconda3\envs\pytorch\python.exe D:/PythonStudySpace/yolov4-pytorch/FPS_test.py
Loading weights into state dict...
Finished!
model_data/yolo4_weights.pth model, anchors, and classes loaded.
2.951487748622894 seconds, 0.3388121805576121FPS, @batch_size 1
Process finished with exit code 0
D:\ProgramData\Anaconda3\envs\pytorch\python.exe D:/PythonStudySpace/yolov4-pytorch/kmeans_for_anchors.py
acc:87.22%
[[ 15.15428571 73.9375 ]
[ 16.04571429 140.96875 ]
[ 22.88 85.71875 ]
[ 23.32571429 40.828125 ]
[ 54.08 49.359375 ]
[ 66.85714286 89.78125 ]
[ 67.74857143 93.84375 ]
[ 68.64 85.71875 ]
[ 71.90857143 72.71875 ]]
(pytorch) ubuntu-gpu@UbuntuGPU:~/wangmiao/yolov4-pytorch-master/yolov4-pytorch-master$ python FPS_test.py
Loading weights into state dict...
Finished!
logs/Epoch70-Total_Loss0.9002-Val_Loss2.2391.pth model, anchors, and classes loaded.
0.030595953464508056 seconds, 32.68406069319006FPS, @batch_size 1
('VOCdevkit/VOC2007/ImageSets/Main/test.txt：352
./VOCdevkit/VOC2007/JPEGImages/testImages/"+image_id+".jpg":353个
统计某文件夹下文件的个数
train and val size 1052
traub suze 1052：353
统计某文件夹下目录的个数
ls -l |grep "^ｄ"|wc -l
统计文件夹下文件的个数，包括子文件夹里的
ls -lR|grep "^-"|wc -l
~/wangmiao/yolov4-pytorch-master/VOCdevkit/VOC2007/JPEGImages$
ls -l |grep "^-"|wc -l
：825
ubuntu-gpu@UbuntuGPU:~
/wangmiao/yolov4-pytorch-master/VOCdevkit/VOC2007/JPEGImages$
ls -l |grep "^ｄ"|wc -l
0
ubuntu-gpu@UbuntuGPU:~
/wangmiao/yolov4-pytorch-master/VOCdevkit/VOC2007/JPEGImages$
ls -lR|grep "^-"|wc -l
1178
ubuntu-gpu@UbuntuGPU:
~/wangmiao/yolov4-pytorch-master/VOCdevkit/VOC2007/JPEGImages/testImages$
ls -lR|grep "^-"|wc -l
353
ubuntu-gpu@UbuntuGPU:
~/wangmiao/yolov4-pytorch-master/VOCdevkit/VOC2007/Annotations$
ls -l |grep "^-"|wc -l
826
ubuntu-gpu@UbuntuGPU
:~/wangmiao/yolov4-pytorch-master/VOCdevkit/VOC2007/Annotations/Annotations$
ls -l |grep "^-"|wc -l
353
ubuntu-gpu@UbuntuGPU:
~/wangmiao/yolov4-pytorch-master/VOCdevkit/VOC2007/Annotations/Annotations1$
ls -l |grep "^-"|wc -l
352
xmlfilepath=r'./Annotations'
saveBasePath=r"./ImageSets/Main/"
25 ftrainval = open(os.path.join(saveBasePath,'trainval.txt'), 'w')
26 #ftest = open(os.path.join(saveBasePath,'test.txt'), 'w')
27 ftrain = open(os.path.join(saveBasePath,'train.txt'), 'w')
28 fval = open(os.path.join(saveBasePath,'val.txt'), 'w')
VOCdevkit/VOC2007/ImageSets/Main/test.txt：352
('VOCdevkit/VOC2007/ImageSets/Main/train.txt：741
——————————————————————————————————
2.此处对应的图像
通过kmeans_for_anchors可以得到yolo_anchors.txt
./VOCdevkit/VOC2007/JPEGImages/testImages/"+image_id+".jpg
python get_dr_txt.py 得到./input/detection-results
./input/images-optional
3.VOCdevkit/VOC2007/ImageSets/Main/train.txt由此
VOCdevkit/VOC2007/JPEGImages
VOCdevkit/VOC2007/Annotations/
python get_get_txt.py 得到./input/ground-truth
4.VOCdevkit/VOC2007/ImageSets/Main/
test.txt：352、
val.txt：0、
train.txt：741、
trainval.txt：741、
二、通过k-means得到聚类anchors
原来的anchors：
1 12, 16, 19, 36, 40, 28, 36, 75, 76, 55, 72, 146, 142, 110, 192, 243, 459, 401
通过kmeans_for_anchor.py ->yolo_anchors.txt
1 15,73, 16,140, 22,85, 23,40, 54,49, 62,89, 67,92, 68,86, 72,73
新的聚类和原始的有一些不一样，那么我再跑一下实验做一下对比
就是把训练好的权重加到yolo.py这个里面，然后再去改一下这个权重，然后去做预测和map的计算。
Epoch99-Total_Loss0.8858-Val_Loss1.5478.pth
这个loss比较低，效果比较好，拿这个做一下预测
为了保留上一次训练留下的input，使用mv input input1，将其重命名为input1
然后执行上面的操作，使用python get_map.py去生成map的结果一直报错
我发现可能是input生成的问题，因此我重新改变文件路径并重新运行
这次get_gt_txt.py有741
而get_dr_txt.py是352
——————————2021.4.18——————————————
新的input：
detection-results：741
ground-truth：352
images-optional：352
原来的input：
76 yolo = mAP_Yolo()
77 image_ids = open('VOCdevkit/VOC2007/ImageSets/Main/test.txt').read().strip().split()
78
79 if not os.path.exists("./input"):
80 os.makedirs("./input")
81 if not os.path.exists("./input/detection-results"):
82 os.makedirs("./input/detection-results")
83 if not os.path.exists("./input/images-optional"):
84 os.makedirs("./input/images-optional")
85
86
87 for image_id in tqdm(image_ids):
88 image_path = "./VOCdevkit/VOC2007/JPEGImages/testImages/"+image_id+".jpg"
89 image = Image.open(image_path)
90 # 开启后在之后计算mAP可以可视化
91 image.save("./input/images-optional/"+image_id+".jpg")
92 yolo.detect_image(image_id,image)
detection-results：1092
ground-truth：1092
images-optional：1092
get_dr_txt.py
f = open("./input/detection-results/"+image_id+".txt","w")
————————————————p—————————————————
实验结果：
0.00% = Lockingplate AP || score_threhold=0.5 : F1=nan ; Recall=0.00% ; Precision=0.00%
D:/PythonStudySpace/yolov4-pytorch/get_map.py:706: RuntimeWarning: invalid value encountered in true_divide
F1 = np.array(rec)*np.array(prec)/(np.array(prec)+np.array(rec))*2
0.33% = Upperlever AP || score_threhold=0.5 : F1=0.02 ; Recall=33.33% ; Precision=0.82%
D:/PythonStudySpace/yolov4-pytorch/get_map.py:706: RuntimeWarning: invalid value encountered in true_divide
F1 = np.array(rec)*np.array(prec)/(np.array(prec)+np.array(rec))*2
74.80% = lockingplate AP || score_threhold=0.5 : F1=0.72 ; Recall=67.28% ; Precision=77.22%
92.82% = normal AP || score_threhold=0.5 : F1=0.90 ; Recall=96.46% ; Precision=85.18%
85.47% = truncatedplugdoorhandle AP || score_threhold=0.5 : F1=0.87 ; Recall=84.85% ; Precision=88.89%
85.04% = upperlever AP || score_threhold=0.5 : F1=0.85 ; Recall=85.25% ; Precision=85.25%
mAP = 48.35%
72.63% = lockingplate AP || score_threhold=0.5 : F1=0.68 ; Recall=78.41% ; Precision=60.00%
86.20% = normal AP || score_threhold=0.5 : F1=0.88 ; Recall=88.99% ; Precision=86.87%
69.40% = truncatedplugdoorhandle AP || score_threhold=0.5 : F1=0.74 ; Recall=77.78% ; Precision=70.00%
55.58% = upperlever AP || score_threhold=0.5 : F1=0.49 ; Recall=51.35% ; Precision=46.34%
mAP = 70.95%
test：601
周日晚上我重新检查了test数据集的txt，并打开文档查验过，都是小写的，标签文件并没有使用大写
truncated plug door handle
upper lever
locking plate
normal
这四种标签类型分别代表了截断塞门手把故障、上拉杆故障、紧锁板故障、以及所有正常的部分
那么导致map计算最后的结果多出来的大写的种类又发生在哪里呢？
1.test数据集已经检查过了
2.标签文件也已经检查过了
input文件夹分为
–input
--------detections-results：存放txt格式的标签文件，1-1179个
----------------------------- 举例：truncated plug door handle 0.9873 85 524 258 634
--------ground-truth：存放txt格式的标签文件，1-1179个
----------------------------- 举例：truncated plug door handle 989 377 1169 503
--------images-optional:存放jpg格式的图像，一共有1179张
我检查了一遍这个文档，都没有大写的类型，那最后结果大写的到底哪里来的呢？
通过检查results这个文档，我发现了
Lockingplate: 1，errors:663 服务器上已修改
Upperlever: 3: ,errors:97、181、488 服务器上已修改
lockingplate: 272
normal: 989
truncatedpIugdoorhandIe: 1,errors:33----->直接改这里
truncatedplugdoorhandle: 66
upperlever: 122
问题处在ground-truth
ctrl+shift+f 全局搜索字符 还可以设置区分大小写
ctrl+shift+n 全局搜索文件名
打包成tar.gz格式压缩包
tar -zcvf renwolesshel.tar.gz /renwolesshel
解压tar.gz格式压缩包
tar zxvf renwolesshel.tar.gz
打包成tar.bz2格式压缩包
tar -jcvf renwolesshel.tar.bz2 /renwolesshel
解压tar.bz2格式的压缩包
tar jxvf renwolesshel.tar.bz2
压缩成zip格式
zip -q -r renwolesshel.zip renwolesshel/
解压zip格式的压缩包
unzip renwolesshel.zip
1，python get_dr_txt.py
2，python get_gt_txt.py
3，mv input.tar.gz ./input3.tar.gz
4，tar -zcvf input.tar.gz ./input
zip -r input.zip ./input
5， sz input.zip
detection-results
现在又开始出现问题了，以后每一次都要记录下来
去检查一下
image_ids = open('VOCdevkit/VOC2007/ImageSets/Main/test.txt').read().strip().split()
VOCdevkit/VOC2007/ImageSets/Main/test.txt：352
('VOCdevkit/VOC2007/ImageSets/Main/train.txt：741
：
if not os.path.exists("./input"):
os.makedirs("./input")
if not os.path.exists("input/ground-truth"):
os.makedirs("input/ground-truth")
if not os.path.exists("./input/images-optional"):
os.makedirs("./input/images-optional")
"./VOCdevkit/VOC2007/JPEGImages/"+image_id+".jpg"
————————————————2020.5.5—————————————————
annotations：1169
-------------------------
train and val size 1052
traub suze 1052
Epoch100-Total_Loss0.6700-Val_Loss2.3610.pth
Epoch100-Total_Loss0.6705-Val_Loss2.1259.pth
Epoch100-Total_Loss0.6740-Val_Loss1.6421.pth
Epoch100-Total_Loss0.6782-Val_Loss2.0120.pth
Epoch100-Total_Loss0.6836-Val_Loss1.5088.pth
Epoch100-Total_Loss0.6877-Val_Loss1.8490.pth
Epoch100-Total_Loss0.7128-Val_Loss2.1981.pth
Epoch100-Total_Loss0.8817-Val_Loss1.4321.pth
-----------------kmeans-----------------------------
Epoch100-Total_Loss0.6700-Val_Loss2.3610.pth
Epoch100-Total_Loss0.6705-Val_Loss2.1259.pth
Epoch100-Total_Loss0.6740-Val_Loss1.6421.pth
Epoch100-Total_Loss0.6782-Val_Loss2.0120.pth
Epoch100-Total_Loss0.6836-Val_Loss1.5088.pth
Epoch100-Total_Loss0.6877-Val_Loss1.8490.pth
Epoch100-Total_Loss0.7128-Val_Loss2.1981.pth
Epoch100-Total_Loss0.7498-Val_Loss1.7838.pth <------
Epoch100-Total_Loss0.8817-Val_Loss1.4321.pth
————————————————2021.5.6—————————————————
Epoch100-Total_Loss0.7498-Val_Loss1.7838.pth
————————————————2021.5.12————————————————————
数据集:1179（原始数据集图像数量）-3（图像无法识别的图像数量）=1176
训练数据集：823张图片
测试数据集：353张图片