File size: 10,359 Bytes

e26e560

开展工作：运用swin-transformer-object-detection进行目标检测任务
第一步:下载swin-transformer-object-detection
第二步：配置环境（最重要）
需要：
cuda 10.2.89
cudnn 7.6.5
python 3.8
vs 2019
pytorch 1.8.0
mmcv 1.3.17
mmdet 2.20.0（最终变成了2.11.0）
命令：
conda install cudatoolkit=10.2 -c https://mirrors.ustc.edu.cn/anaconda/pkgs/free/win-64/
conda install cudnn=7.6.5 -c https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main/win-64/
conda install pytorch==1.8.0 torchvision==0.9.0 torchaudio==0.8.0


conda install pytorch==1.8.1 torchvision==0.9.1 torchaudio==0.8.1 cudatoolkit=10.2 -c pytorch
conda install pytorch==1.8.0 torchvision==0.9.0 torchaudio==0.8.0 cudatoolkit=10.2
conda install pytorch==1.8.0 torchvision==0.9.0 torchaudio==0.8.0 cudatoolkit=11.0 -c pytorch -c conda-forge  可以安装

 CUDA_PATH CUDA_HOME envs ls env
conda install pytorch torchvision cudatoolkit=10.2 -c pytorch
pyt1.8.1 + torchvision0.9.1 + cuda10.2
https://download.openmmlab.com/mmcv/dist/index.html
pip install mmcv-full==1.3.17 -f https://download.openmmlab.com/mmcv/dist/cu102/torch1.8.0/index.html
swin-T-pytorch18
conda create -n swin-T-pytorch18 --clone pytorch18

installing mmdet from https://github.com/open-mmlab/mmdetection.git.
Cloning into 'C:\Users\zjf\AppData\Local\Temp\tmptf6pnzkx\mmdetection'.
pip install -r requirements.txt -i http://mirrors.aliyun.com/pypi/simple/ --trusted-host
pip install mmcv-full==1.3.17 -f https://download.openmmlab.com/mmcv/dist/cu102/torch1.8.0/index.html
复制前：33.5GB
复制后：
MMDetection V2.24.1 Release



conda list | grep cuda，然后就会获得虚拟环境下安装的版本
conda search cudnn --info  查看所有cuda版本和对应的cudnn版本

查询显卡算力
进入目录 deviceQuery.exe
>> cd D:\360Downloads\CUDA\NVIDIA GPU Computing Toolkit\CUDA\v10.1\extras\demo_suite
 
运行 deviceQuery.exe
>> ./deviceQuery.exe
我的显卡算力为7.5
Device 0: "NVIDIA GeForce GTX 1660 Ti"
  CUDA Driver Version / Runtime Version          11.6 / 11.0
  CUDA Capability Major/Minor version number:    7.5
、deviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 11.6, CUDA Runtime Version = 11.0, NumDevs = 1, Device0 = NVIDIA GeForce GTX 1660 Ti

E:\CUDA\v11.0\bin>nvcc -V
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2020 NVIDIA Corporation
Built on Wed_Jul_22_19:09:35_Pacific_Daylight_Time_2020
Cuda compilation tools, release 11.0, V11.0.221
Build cuda_11.0_bu.relgpu_drvr445TC445_37.28845127_0




True
10.2
0
1
NVIDIA GeForce GTX 1660 Ti
1.8.0
10.2
7605



运行demo
图片：
python demo/image_demo.py demo/demo.jpg configs/swin/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco.py mask_rcnn_swin_tiny_patch4_window7.pth

# 变成了纯目标检测，配置文件已经改变，所以之前的权重文件已经失效
python demo/image_demo.py demo/test.png configs/swin/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco.py mask_rcnn_swin_tiny_patch4_window7.pth


python demo/image_demo.py demo/demo.jpg configs/swin/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco.py mask_rcnn_swin_tiny_patch4_window7.pth


视频：
保存：
 python demo/video_demo.py demo/demo.mp4 configs/swin/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco.py mask_rcnn_swin_tiny_patch4_window7.pth --out demo\1.mp4

展示：
python demo/video_demo.py demo/demo.mp4 configs/swin/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco.py mask_rcnn_swin_tiny_patch4_window7.pth --show

python demo/video_demo.py demo/demo.mp4 configs/swin/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco.py latest.pth --show

 python demo/image_demo.py demo/test.png configs/swin/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco.py latest.pth



我的推理测试命令：
 python demo/image_demo.py demo/6.jpg configs/swin/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco.py checkpoints/epoch_50.pth

python demo/image_demo.py demo/demo.jpg configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py checkpoints/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth



2022年5月24日20：40，历时两天，终于安装环境成功，版本为1.8.0 10.2.89 7.6.5

coco数据集是json格式
voc数据集是xml格式

初步决定先做mask r cnn做实例分割
再做faster rcnn做纯目标检测

实例分割：相当于带有mask的目标检测

实例分割
目标检测和语义分割的结合，在图像中将目标检测出来（目标检测），然后对目标每个像素打上标签（语义分割），不需要将图像中所有像素打上标签

需要COCO格式


CLASSES = ['ignored regions', 'pedestrian', 'people', 'bicycle', 'car', 'van',
              'truck', 'tricycle', 'awning-tricycle', 'bus', 'motor', 'others']

三个py文件
修改Swin-Transformer-Object-Detection/mmdet/datasets/coco.py的CLASSES
修改Swin-Transformer-Object-Detection/mmdet/core/evaluation/class_names.py 的coco_classes
修改路径/configs/base/models/mask_rcnn_swin_fpn.py中的num_classes，也有两处大概在第54行和73行，修改为自己数据集的类别数量
————————————————
原文链接：https://blog.csdn.net/qq_41964545/article/details/123140485


4.修改训练参数
路径/configs/swin/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x(1x)_coco.py中：

第3行’…/base/datasets/coco_instance.py’修改为’…/base/datasets/coco_detection.py’
第69行的max_epochs按需修改
第31行的samples_per_gpu表示batch size大小，太大会内存溢出
第32行的workers_per_gpu表示每个GPU对应线程数，2、4、6、8按需修改
————————————————
原文链接：https://blog.csdn.net/qq_36622589/article/details/124355564



制作数据集：
将数据集从txt转换成xml，再转换成json


checkpoint中包含了需要加载模型的目录



这里以mask_rcnn_swin_fpn作为backbone为例，数据格式为coco数据格式，其他模型的修改等同


训练命令
python tools/train.py configs/swin/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco.py

python tools/train.py  configs/swin/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco.py --cfg-options model.pretrained='checkpoints/latest.pth'

python tools/train.py  configs/swin/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco.py --cfg-options model.pretrained='checkpoints/mask_rcnn_swin_tiny_patch4_window7.pth'


python tools/train.py  configs/swin/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco.py --resume-from  model.pretrained='checkpoints/latest.pth'


python tools/train.py  configs/swin/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco.py --resume-from  model.pretrained='checkpoints/epoch_50.pth'  这不行

python tools/train.py  configs/swin/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco.py --resume-from 'checkpoints/epoch_75.pth'  这可以 断点续传



测试命令:

python demo/image_demo.py demo/000019.jpg configs\swin\mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco.py work_dirs/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco/latest.pth


python demo/image_demo.py testfiles/img1.jpg 
configs/swin/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_1x_coco.py 
work_dirs/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_1x_coco/latest.pth --score-thr 0.5

python demo/video_demo.py demo/demo.mp4 configs/swin/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco.py work_dirs/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco/latest.pth --score-thr 0.5 --show


3）性能统计

python tools/test.py configs/swin/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco.py work_dirs/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco/latest.pth --eval bbox

4）日志分析

python tools/analysis_tools/analyze_logs.py plot_curve work_dirs/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco/20220527_092210.log.json


id字段：指的是这个annotation的一个id
image_id：等同于前面image字段里面的id。
category_id：类别id
segmentation：用于分割         我的数据集中没有这一项,所以完成不了实例分割,只能进行目标检测
area：标注区域面积
bbox：标注框，左上角坐标 标注框宽和高
iscrowd：决定是RLE格式还是polygon格式。

参考文献
https://zhuanlan.zhihu.com/p/451816231
https://github.com/SwinTransformer/Swin-Transformer-Object-Detection/issues/113
https://blog.csdn.net/qq_41964545/article/details/123140485
# 去除实例分割,仅进行目标检测
https://github.com/SwinTransformer/Swin-Transformer-Object-Detection/issues/50
https://github.com/SwinTransformer/Swin-Transformer-Object-Detection/issues/25 有效
# 训练目标检测
https://blog.csdn.net/qq_36622589/article/details/124355564?spm=1001.2101.3001.6650.1&utm_medium=distribute.pc_relevant.none-task-blog-2%7Edefault%7ECTRLIST%7ERate-1-124355564-blog-123140485.pc_relevant_paycolumn_v3&depth_1-utm_source=distribute.pc_relevant.none-task-blog-2%7Edefault%7ECTRLIST%7ERate-1-124355564-blog-123140485.pc_relevant_paycolumn_v3&utm_relevant_index=2

将visdrone数据集转化为coco格式并在mmdetection上训练,附上转好的json文件
https://blog.csdn.net/S5242/article/details/121114907
https://github.com/VisDrone/VisDrone-Dataset
http://aiskyeye.com/download/object-detection-2/


https://beyonderwei.com/2022/03/23/Swin-Transformer%E7%9B%AE%E6%A0%87%E6%A3%80%E6%B5%8B4%E2%80%94%E2%80%94%E8%AE%AD%E7%BB%83%E8%87%AA%E5%B7%B1%E6%95%B0%E6%8D%AE%E9%9B%86/

https://blog.csdn.net/CSDN_X_W/article/details/123845728


https://blog.csdn.net/m0_37605642/article/details/117932717

https://blog.csdn.net/qq_42138662/article/details/109227007
coco数据集介绍
https://blog.csdn.net/qq_44554428/article/details/122597358


Ze Liu Yutong Lin Yue Cao Han Hu Yixuan Wei Zheng Zhang Stephen Lin Baining Guo
Swin Transformer : Hierarchical Vision Transformer using Shifted Windows. ICCV 2021

swin_t心路之旅
read://https_blog.51cto.com/?url=https%3A%2F%2Fblog.51cto.com%2Fu_13565704%2F5105836%3Fb%3Dtotalstatistic


https://wenku.baidu.com/view/9aa300a3bfeb19e8b8f67c1cfad6195f312be8cb.html