Configure MMdetection environment and train

配置MMdetection环境（以VFNet为例）

It is recommended to use the official website method first,链接放在下面了！

https://github.com/open-mmlab/mmdetection/blob/master/docs/get_started.md

1、先创建一个conda环境并激活它

conda create -n open-mmlab python=3.7 -y
conda activate open-mmlab

2、安装PyTorch 和 torchvision,
根据cuda和pytorch版本安装（我的版本是cuda11.0和torch1.7.0）

conda install pytorch torchvision -c pytorch

conda install pytorch cudatoolkit=11.0 torchvision -c pytorch

3、安装mmcv-full
根据cuda和pytorch版本安装（我的版本是cuda11.0和torch1.7.0）.mmcvIf the version configuration is incorrect, the model cannot be trained normally,Be sure to use the correct version.

pip install mmcv-full -f https://download.openmmlab.com/mmcv/dist/cu110/torch1.7.0/index.html

4、安装pycocotools
Although the official website does not require installation,But I am getting error missingpycocotools.

pip install git+https://github.com/philferriere/cocoapi.git#subdirectory=PythonAPI

或者https://github.com/philferriere/cocoapi下载源码,并进行解压.切换到 cocoapi\PythonAPI目录.运行：

python setup.py build_ext install

然后cdto your own path,安装mmdetection
5、mmdetection安装和编译

git clone https://github.com/open-mmlab/mmdetection.git
cd mmdetection
pip install -r requirements/build.txt
pip install -v -e.  # or "python setup.py develop".

到这里为止,环境的配置就完成了.Start preparing the dataset
6、使用自己的coco格式数据集
All officially provided code is used by defaultcoco格式的数据集,I also use it herecoco格式的数据集.Various types of data transfercocoSee format script：

https://github.com/spytensor/prepare_detection_dataset

mmdetection
├── mmdet
├── tools
├── configs
├── data
│ ├── coco
│ │ ├── annotations
│ │ ├── train2017
│ │ ├── val2017
│ │ ├── test2017

7、Some parameters need to be changed before training
1）定义数据种类,The place that needs to be modified is heremmdetection/mmdet/datasets/coco.py.把CLASSES的那个tuple改为自己数据集对应的种类tuple即可.

CLASSES = ('liner', 'bulk carrier', 'container ship', 'other ship')

2）接着在mmdetection/mmdet/core/evaluation/class_names.py修改coco_classes数据集类别,This is related to the backtestThe category name displayed in the resulting graph when

3）修改configs/vfnet/vfnet_x101_64x4d_fpn_mstrain_2x_coco.py中的model字典中的num_classes、data字典中的img_scale和optimizer中的lr(学习率).例如：

num_classes=4,#类别数
optimizer = dict(type='SGD', lr=0.0025, momentum=0.9, weight_decay=0.0001)

当gpu数量为8时,lr=0.02;当gpu数量为4时,lr=0.01;我只要一个gpu,所以设置lr=0.0025

训练完之后work_dirs文件夹中会保存下训练过程中的log日志文件、每个epoch的pth文件（这个文件将会用于后面的test测试）

8、开始训练
单个GPU（Not distributed）:

#Train the model from the start
python tools/train.py configs/vfnet/vfnet_x101_64x4d_fpn_mstrain_2x_coco.py
#Continue training from the breakpoint weights file
python tools/train.py configs/vfnet/vfnet_x101_64x4d_fpn_mstrain_2x_coco.py --resume-from="work_dirs/yy.pth"

resume_from和load_from之间的区别： resume_from加载模型权重和优化器状态,​​并且epoch也从指定的检查点继承.它通常用于恢复意外中断的训练过程. load_from仅加载模型权重,训练时期从0开始.通常用于微调.

9、测试
（1）If you just want to see the effect and don't do quantitative indicator analysis,Can run the previous onedemo.py文件,But change itcheckpoint_file的地址路径,Use what we ran out of the previous stepwork_dirs下的pth文件.例如checkpoint_file = 'work_dirs/epoch_100.pth

（2）使用testcommand to perform a test to evaluate some parameters

 python tools/test.py configs/vfnet/vfnet_x101_64x4d_fpn_mstrain_2x_coco.py work_dirs/latest.pth --out ./result/result_100.pkl --eval bbox

Put a picture of the result of my training（The dataset used is its own,single class 2万多张）
VFNet X-101-64x4d (mAP 91.7%)

10、报错和解决
1）修改config里面的num_calss之后报错：
重新编译

 python setup.py install

要是还不行,先卸载mmdet,再重新编译

pip uninstall mmdet
pip install -r requirements/build.txt
pip install -v -e .  # or "python setup.py develop".

Modifications will be added later！