Faster-ILOD、maskrcnn_benchmark训练自己的voc数据集及问题汇总

一、训练自己标注的VOC数据集

使用Faster-ILOD训练自己标注的数据集，数据集已经转换为VOC格式
之前已经可以顺利训练VOC2007数据集，可参考
Faster-ILOD、maskrcnn_benchmark安装过程及遇到问题
主要修改以下文件：

1.maskrcnn_benchmark/config/paths_catalog.py

对应路径修改为自己数据集所在的路径

        "voc_2007_train": {
    
            "data_dir": "/data/taxi_data/VOC2007_all",
            "split": "train"
        },
        "voc_2007_val": {
    
            "data_dir": "/data/taxi_data/VOC2007_all",
            "split": "val"
        },
        "voc_2007_test": {
    
            "data_dir": "/data/taxi_data/VOC2007_all",
            "split": "test"
        },

2.maskrcnn_benchmark/data/voc.py

修改为数据集中包含类别 我主要进行出租车和私家车的识别

    # CLASSES = ("__background__ ", "aeroplane", "bicycle", "bird", "boat", "bottle", "bus", "car", "cat", "chair", "cow",
    # "diningtable", "dog", "horse", "motorbike", "person", "pottedplant", "sheep", "sofa", "train", "tvmonitor")
    CLASSES = ("__background__ ", "Taxi","Other Car")

3.修改configs/e2e_faster_rcnn_R_50_C4_1x.yaml

修改类别数量等，我一次把两个类都训练了。

NUM_CLASSES: 3 # total classes 
NAME_OLD_CLASSES: []
#NAME_NEW_CLASSES: ["aeroplane", "bicycle", "bird", "boat", "bottle", "bus", "car", "cat", "chair", "cow", "diningtable", "dog",
# "horse", "motorbike", "person","pottedplant", "sheep", "sofa", "train","tvmonitor"]
# NAME_EXCLUDED_CLASSES: [ ]
NAME_NEW_CLASSES: ["Taxi","Other Car"]
NAME_EXCLUDED_CLASSES: []

4.运行

python tools/train_first_step.py --config-file="./configs/e2e_faster_rcnn_R_50_C4_1x.yaml"

二、遇到问题

1.没有对应的Taxi_train.txt 等文件

因为是做增量学习，会针对每一个类别进行处理，原始下载的voc文件中是包含每一类别对应的train、test等txt文件，但对自己标注数据集转换为voc格式的数据集时中转换为train.txt,test.txt,trainval.txt以及val.txt。没有对应的txt文件，因此要生成对应的文件。

观察元数据集中每一类别对应的txt文件，每一行对应两列，一列为图片名称，另一列是一位数字，应该值为1.-1.0.
-1代表该类不存在此图片中
1位存在此图片中
0为困难样本
（个人觉得是这样，不对的欢迎指正）

根据xml文件和train.txt,test.txt,val.txt进行对应类别的转换。代码如下：
我生成对应文件时只将存在此类别的图片名称写入，在数据读取的时候报错就改了一下代码，可以修改一下代码，与原数据集保持一致。

names = locals()
dir_path = "D:/achenf/data/jiaotong_data/1taxi_train_data/VOC/VOC2007/"
# 读取train.txt/test.txt/val.txt文件
f = open("D:/achenf/data/jiaotong_data/1taxi_train_data/VOC/VOC2007/ImageSets/Main/train.txt", 'r')
# 包含类别
classes = ['Taxi','Other Car']
# 写入文档名称
txt_path = dir_path+"by_classes/"+"Other Car_train.txt"
fp_w = open(txt_path, 'w')
for line in f.readlines():
    if not line:
        break
    n= line[:-1]
    xmlpath=dir_path+"Annotations/"+n+'.xml'
    fp = open(xmlpath)

    xmllines = fp.readlines()
    ind_start = []
    ind_end = []
    lines_id_start = xmllines[:]
    lines_id_end = xmllines[:]
    # 修改对应类别Other Car/Taxi
    classes1 = ' <name>Other Car</name>\n'
    while " <object>\n" in lines_id_start:
        a = lines_id_start.index(" <object>\n")
        ind_start.append(a)
        lines_id_start[a] = "delete"

    while " </object>\n" in lines_id_end:
        b = lines_id_end.index(" </object>\n")
        ind_end.append(b)
        lines_id_end[b] = "delete"

    # names中存放所有的object块
    i = 0
    for k in range(0, len(ind_start)):
        names['block%d' % k] = []
        for j in range(0, len(classes)):
            if classes[j] in xmllines[ind_start[i] + 1]:
                a = ind_start[i]
                for o in range(ind_end[i] - ind_start[i] + 1):
                    names['block%d' % k].append(xmllines[a + o])
                break
        i += 1
    # 在给定的类中搜索，若存在则，写入object块信息
    a = 0
    if len(ind_start)>0:
        # xml头
        string_start = xmllines[0:ind_start[0]]
        # xml尾
        string_end = [xmllines[len(xmllines) - 1]]
        flag = False
        for k in range(0, len(ind_start)):
            if classes1 in names['block%d' % k]:
                flag = True
                a += 1
                string_start += names['block%d' % k]
        string_start += string_end
        # 如果存在写入
        if flag:
            fp_w.write(n+" "+"1"+'\n')
    fp.close()
fp_w.close()

得到以下文件：

2.x[2] == ‘0’ 超出下标

对每一行就行分割时，如果值为1，源数据集中txt文件分解为[‘000000’,‘’,‘1’],我生成的文件分解为[‘000000’,‘1’],因为没有困难样本，直接将这几行注释

    for i in range(len(buff)):
        x = buff[i]
        x = x.split(' ')
        if x[1] == '-1':
            pass
        # elif x[2] == '0': # include difficult level object
        # if self.is_train:
        # pass
        # else:
        # img_ids_per_category.append(x[0])
        # self.ids.append(x[0])
        else:
            img_ids_per_category.append(x[0])
            self.ids.append(x[0])

以及本人的xml文件中也不存在difficult的标签，所以遇到此行代码difficult = int(obj.find("difficult").text) == 1 也会报错，所以直接将difficult赋值为False

for obj in target.iter("object"):

     difficult = False
     # difficult = int(obj.find("difficult").text) == 1
     if not self.keep_difficult and difficult:
         continue

3.ValueError: invalid literal for int() with base 10: ‘0.0’

报错出在此行bndbox = tuple(map(lambda x: x - TO_REMOVE, list(map(int, box))))，应该是str直接转为int的问题，于是我们先将其转换为float，在转换为int，修改如下：bndbox = tuple(map(lambda x: x - TO_REMOVE, list(map(int, map(float,box)))))
就ok了
学习到一个新知识点，map（）函数可以实现批量数据类型转换。

可参考： Python3中用map()批量地转换数据类型，例如str转float

4.Pytorch报错 CUDA error: device-side assert triggered

参考：https://blog.csdn.net/veritasalice/article/details/111917185