Faster_Rcnn 训练自定义数据集

Faster Rcnn 训练自定义数据集

 

目录

• Faster Rcnn 训练自定义数据集
• 运行原代码
  • • 下载代码
    • 创建文件夹
    • 下载预训练模型
    • 下载数据集
    • 配置环境
    • 更换coco api
    • 进行编译
    • 训练
    • 测试
    • 运行demo（画框程序）
• 训练自定义数据集
  • • 格式转换
    • 替换原VOC2007数据集
    • 更改类别
    • 解决标定区域溢出

 

运行原代码

下载代码

使用jwyang的pytorch框架代码，由于主分支支持的pytorch版本为0.4，年代久远，可能出现许多问题，这边选择pytorch-1.0分支。

创建文件夹

创建文件夹data和子文件夹pretrained_model、VOCdevkit2007，用于存放预训练模型和数据集

cd faster-rcnn.pytorch-pytorch-1.0 && mkdir data
cd data && mkdir pretrained_model VOCdevkit2007

下载预训练模型

下载官方在caffe上的训练的预训练模型，VGG和ResNet选一即可，本文以ResNet为例，VT Server服务商国内可以裸连下载。下载完成后，将该模型放置到./data/pretrained_model

• VGG Dropbox VT Server
• ResNet101 Dropbox VT Server

下载数据集

下载VOC数据集，解压完后放置到./data/VOCdevkit2007

wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCdevkit_08-Jun-2007.tar

tar xvf VOCtrainval_06-Nov-2007.tar
tar xvf VOCtest_06-Nov-2007.tar
tar xvf VOCdevkit_08-Jun-2007.tar

配置环境

创建虚拟环境，完成环境配置，pytorch 版本1.0.0，torchvision版本0.2.1，scipy版本1.2.1，其余支持按requirements.txt要求默认安装即可，pip库时建议使用豆瓣源。

conda create -n fr python=3.6
conda activate fr
pip install pytorch==1.0 torchvision==0.2.1 scipy==1.2.1 -i http://pypi.douban.com/simple/ --trusted-host pypi.douban.com

更换coco api

使用最新的官方coco api替换原代码的coco api

git clone https://github.com/pdollar/coco.git
cd cocoapi/PythonAPI
make

make时注意在创建的fr环境下进行，否则可能导致失败，make完成后，用~/cocoapi/PythonAPI/pycocotools 替换~/faster-rcnn.pytorch-pytorch-1.0/lib/pycocotools

进行编译

cd ~/faster-rcnn.pytorch-pytorch-1.0/lib
python setup.py build develop

训练

进行训练，使用GPU0，网络采用ResNet101，batch size设置为4（8G显存下可行），线程设置为2，运行100 epochs

CUDA_VISIBLE_DEVICES=0 python trainval_net.py --dataset pascal_voc --net res101 --bs 4 --nw 2 --epochs 100 --cuda

测试

使用GPU0，网络采用ResNet101，载入faster_rcnn_1_20_470.pth模型

CUDA_VISIBLE_DEVICES=0 python test_net.py --dataset pascal_voc --net res101 --checksession 1 --checkepoch 20 --checkpoint 470 --cuda

运行demo（画框程序）

在./images存放需要画框的图片，输入指令运行，参数与上类似

CUDA_VISIBLE_DEVICES=0 	python demo.py --net res101 \
               			--checksession 1 --checkepoch 100 --checkpoint 470 \
               			--cuda --load_dir ./models

训练自定义数据集

格式转换

所要做的就是将自己的数据集替换VOC2007，由于我的数据集是yolo格式，因此要先进行转换

from xml.dom.minidom import Document
import os
import cv2
from tqdm import tqdm

# def makexml(txtPath, xmlPath, picPath):  # txt所在文件夹路径，xml文件保存路径，图片所在文件夹路径
def makexml(picPath, txtPath, xmlPath):  # txt所在文件夹路径，xml文件保存路径，图片所在文件夹路径
    """此函数用于将yolo格式txt标注文件转换为voc格式xml标注文件
    """
    dic = {'0': "holothurian",  # 创建字典用来对类型进行转换
           '1': "echinus",  
           '2': 'scallop',
           '3': 'starfish',
           '4': 'waterweeds'
           }
    files = os.listdir(txtPath)
    for i, name in tqdm(enumerate(files)):
        xmlBuilder = Document()
        annotation = xmlBuilder.createElement("annotation")  # 创建annotation标签
        xmlBuilder.appendChild(annotation)
        txtFile = open(txtPath + name)
        txtList = txtFile.readlines()
        img = cv2.imread(picPath + name.split('.')[0] + ".jpg")
        Pheight, Pwidth, Pdepth = img.shape

        folder = xmlBuilder.createElement("folder")  # folder标签
        foldercontent = xmlBuilder.createTextNode("driving_annotation_dataset")
        folder.appendChild(foldercontent)
        annotation.appendChild(folder)  # folder标签结束

        filename = xmlBuilder.createElement("filename")  # filename标签
        filenamecontent = xmlBuilder.createTextNode(name[0:-4] + ".jpg")
        filename.appendChild(filenamecontent)
        annotation.appendChild(filename)  # filename标签结束

        size = xmlBuilder.createElement("size")  # size标签
        width = xmlBuilder.createElement("width")  # size子标签width
        widthcontent = xmlBuilder.createTextNode(str(Pwidth))
        width.appendChild(widthcontent)
        size.appendChild(width)  # size子标签width结束

        height = xmlBuilder.createElement("height")  # size子标签height
        heightcontent = xmlBuilder.createTextNode(str(Pheight))
        height.appendChild(heightcontent)
        size.appendChild(height)  # size子标签height结束

        depth = xmlBuilder.createElement("depth")  # size子标签depth
        depthcontent = xmlBuilder.createTextNode(str(Pdepth))
        depth.appendChild(depthcontent)
        size.appendChild(depth)  # size子标签depth结束

        annotation.appendChild(size)  # size标签结束

        for j in txtList:
            oneline = j.strip().split(" ")
            object = xmlBuilder.createElement("object")  # object 标签
            picname = xmlBuilder.createElement("name")  # name标签
            namecontent = xmlBuilder.createTextNode(dic[oneline[0]])
            picname.appendChild(namecontent)
            object.appendChild(picname)  # name标签结束

            pose = xmlBuilder.createElement("pose")  # pose标签
            posecontent = xmlBuilder.createTextNode("Unspecified")
            pose.appendChild(posecontent)
            object.appendChild(pose)  # pose标签结束

            truncated = xmlBuilder.createElement("truncated")  # truncated标签
            truncatedContent = xmlBuilder.createTextNode("0")
            truncated.appendChild(truncatedContent)
            object.appendChild(truncated)  # truncated标签结束

            difficult = xmlBuilder.createElement("difficult")  # difficult标签
            difficultcontent = xmlBuilder.createTextNode("0")
            difficult.appendChild(difficultcontent)
            object.appendChild(difficult)  # difficult标签结束

            bndbox = xmlBuilder.createElement("bndbox")  # bndbox标签
            xmin = xmlBuilder.createElement("xmin")  # xmin标签
            mathData = int(((float(oneline[1])) * Pwidth + 1) - (float(oneline[3])) * 0.5 * Pwidth)
            xminContent = xmlBuilder.createTextNode(str(mathData))
            xmin.appendChild(xminContent)
            bndbox.appendChild(xmin)  # xmin标签结束

            ymin = xmlBuilder.createElement("ymin")  # ymin标签
            mathData = int(((float(oneline[2])) * Pheight + 1) - (float(oneline[4])) * 0.5 * Pheight)
            yminContent = xmlBuilder.createTextNode(str(mathData))
            ymin.appendChild(yminContent)
            bndbox.appendChild(ymin)  # ymin标签结束

            xmax = xmlBuilder.createElement("xmax")  # xmax标签
            mathData = int(((float(oneline[1])) * Pwidth + 1) + (float(oneline[3])) * 0.5 * Pwidth)
            xmaxContent = xmlBuilder.createTextNode(str(mathData))
            xmax.appendChild(xmaxContent)
            bndbox.appendChild(xmax)  # xmax标签结束

            ymax = xmlBuilder.createElement("ymax")  # ymax标签
            mathData = int(((float(oneline[2])) * Pheight + 1) + (float(oneline[4])) * 0.5 * Pheight)
            ymaxContent = xmlBuilder.createTextNode(str(mathData))
            ymax.appendChild(ymaxContent)
            bndbox.appendChild(ymax)  # ymax标签结束

            object.appendChild(bndbox)  # bndbox标签结束

            annotation.appendChild(object)  # object标签结束

        f = open(xmlPath + name[0:-4] + ".xml", 'w')
        xmlBuilder.writexml(f, indent='\t', newl='\n', addindent='\t', encoding='utf-8')
        f.close()

if __name__ == "__main__":
    picPath = "./pic/train/"  # 图片所在文件夹路径，后面的/一定要带上
    txtPath = "./labels/train/"  # txt所在文件夹路径，后面的/一定要带上
    xmlPath = "./xml/train/"  # xml文件保存路径，后面的/一定要带上
		#运行前请先修改makexml函数中dic
    makexml(picPath, txtPath, xmlPath)

import glob

#图片的地址
train_image_path = r"./pic/train/"
trainvalid_image_path = r"./pic/test/"
test_image_path = r"./pic/test/"
#生成的txt的路径
txt_path = r"./txt/"

def generate_train_and_val(image_path, txt_file):
    with open(txt_file, 'w') as tf:
        for jpg_file in glob.glob(image_path + '*.jpg'):
            jpg_file = jpg_file.split('/')[-1].split('.')[0]
            tf.write(jpg_file + '\n')

generate_train_and_val(train_image_path, txt_path + 'train.txt')
generate_train_and_val(trainvalid_image_path, txt_path + 'trainvalid.txt')
generate_train_and_val(test_image_path, txt_path + 'test.txt')

替换原VOC2007数据集

VOC2007数据集格式

–VOC2007
-------------Annotations
--------------ImagesSet
-------------------------Main
----------------------------trainval.txt
----------------------------train.txt
----------------------------test.txt
-------------JPEGImages

Annotations存放所有的XML文件，JPEGImages存放所有图片,./ImagesSet/Main存放索引TXT文件，将自己的数据集按此替换即可

更改类别

1. 进入./lib/datasets/pascal_voc.py，更改48行self._classes为自己的类别，注意不要更改__background__

# before
self._classes = ('__background__',  # always index 0
                         'aeroplane', 'bicycle', 'bird', 'boat',
                         'bottle', 'bus', 'car', 'cat', 'chair',
                         'cow', 'diningtable', 'dog', 'horse',
                         'motorbike', 'person', 'pottedplant',
                         'sheep', 'sofa', 'train', 'tvmonitor')
#after
self._classes = ('__background__',  # always index 0
                 'holothurian',
                 'echinus',
                 'scallop',
                  'starfish',
                  'waterweeds')

1. 将第243行cls中的.lower()去掉，使标注类别区分大小写

# before
cls = self._class_to_ind[obj.find('name').text.lower().strip()]
 
# after
cls = self._class_to_ind[obj.find('name').text.strip()]

1. 若需要运行demo.py，则需要对其168行pascal_classes的类别进行相应更改

解决标定区域溢出

1. 打开./lib/datasets/imdb.py，找到append_flipped_images函数，进行更改

#before
def append_flipped_images(self):
    num_images = self.num_images
    widths = self._get_widths()
    for i in range(num_images):
      boxes = self.roidb[i]['boxes'].copy()
      oldx1 = boxes[:, 0].copy()
      oldx2 = boxes[:, 2].copy()
      boxes[:, 0] = widths[i] - oldx2 - 1
      boxes[:, 2] = widths[i] - oldx1 - 1
      assert (boxes[:, 2] >= boxes[:, 0]).all()
      entry = {'boxes': boxes,
               'gt_overlaps': self.roidb[i]['gt_overlaps'],
               'gt_classes': self.roidb[i]['gt_classes'],
               'flipped': True}
      self.roidb.append(entry)
    self._image_index = self._image_index * 2

#after
def append_flipped_images(self):
    num_images = self.num_images
    widths = self._get_widths()
    for i in range(num_images):
      boxes = self.roidb[i]['boxes'].copy()
      oldx1 = boxes[:, 0].copy()
      oldx2 = boxes[:, 2].copy()
			#---------add--------------
			for b in range(len(boxes)):
			  if boxes[b][2]< boxes[b][0]:
			    boxes[b][0] = 0
			#---------end--------------
      boxes[:, 0] = widths[i] - oldx2 - 1
      boxes[:, 2] = widths[i] - oldx1 - 1
      assert (boxes[:, 2] >= boxes[:, 0]).all()
      entry = {'boxes': boxes,
               'gt_overlaps': self.roidb[i]['gt_overlaps'],
               'gt_classes': self.roidb[i]['gt_classes'],
               'flipped': True}
      self.roidb.append(entry)
    self._image_index = self._image_index * 2

1. 打开./lib/datasets/pascal_voc.py，找到_load_pascal_annotation，对下面的for循环进行更改

#before
for ix, obj in enumerate(objs):
            bbox = obj.find('bndbox')
            # Make pixel indexes 0-based
            x1 = float(bbox.find('xmin').text) - 1
            y1 = float(bbox.find('ymin').text) - 1
            x2 = float(bbox.find('xmax').text) - 1
            y2 = float(bbox.find('ymax').text) - 1

            diffc = obj.find('difficult')
            difficult = 0 if diffc == None else int(diffc.text)
            ishards[ix] = difficult
#after
for ix, obj in enumerate(objs):
            bbox = obj.find('bndbox')
            # Make pixel indexes 0-based
			#------------change------------------
            x1 = float(bbox.find('xmin').text)
            y1 = float(bbox.find('ymin').text)
            x2 = float(bbox.find('xmax').text)
            y2 = float(bbox.find('ymax').text)
			cls = self._class_to_ind[obj.find('name').text.lower().strip()]
			#--------------end-------------------
            diffc = obj.find('difficult')
            difficult = 0 if diffc == None else int(diffc.text)
            ishards[ix] = difficult

最后请检查是否删除了./data/cache，然后开跑。