/LSNet_supplement

Project material supplement

LSNet_supplement

Project material supplement

LSNet

https://github.com/Duankaiwen/LSNet

安装包

matplotlib

shapely

scipy

torch==1.5.0

torch==0.6.0

（1）安装mmcv, 例如cuda版本 10.0.130

pip3 install mmcv-full -f  https://download.openmmlab.com/mmcv/dist/cu101/torch1.5.0/index.html

（2）安装mmdet

cd /home/yq/work-test/2021/LSNet-main/code

python3 setup.py develop

（3）安装cocoapi

a. 先安装pip3 install cython

b. cd code/cocoapi/pycocotools

c. Python3 setup.py develop

（4）修改

./code/mmdet/ops/conv_ws.py

第25行改成：@CONV_LAYERS.register_module(name='ConvWS', force=True)

第52行改成：@CONV_LAYERS.register_module(name='ConvAWS', force=True)

（5）ImportError: libcudart.so.10.0: cannot open shared object file: No such file or directory

如果遇到错误ImportError: libcudart.so.10.0: cannot open shared object file: No such file or directory

sudo ldconfig /usr/local/cuda-10.0/lib64

（6）修改训练基本设置

路径：code/configs

'./_base_/datasets/coco_lsvr.py', #数据设置

'./_base_/schedules/schedule_1x.py',#模型设置

'./_base_/default_runtime.py'

（7）Generate extreme point annotation from segmentation:

cd code/tools
python gen_coco_lsvr.py
生成instances_lsvr_train2017.json，instances_lsvr_val2017.json，放入路径coco/annotations

（8）训练

detection

./code/tools/dist_train.sh ./code/configs/lsnet/lsnet_bbox_r50_fpn_1x_coco.py 2 --work-dir work_dir/lsnet_bbox_r50_fpn_1x_coco

detection + pose

./code/tools/dist_train.sh ./code/configs/lsnet/lsnet_pose_bbox_r50_fpn_1x_coco.py 2 --work-dir work_dir/lsnet_pose_bbox_r50_fpn_1x_coco

(9)测试

detection

python3 code/tools/test.py code/configs/lsnet/lsnet_bbox_r50_fpn_1x_coco.py work_dir/lsnet_bbox_r50_fpn_1x_coco/epoch_12.pth --show-dir ./results

detection + pose

python3 code/tools/test.py code/configs/lsnet/lsnet_pose_bbox_r50_fpn_1x_coco.py work_dir/lsnet_pose_bbox_r50_fpn_1x_coco/epoch_12.pth --show-dir ./results

可选：使用虚拟环境

用virtualenv包