（slowfast使用方法二） -- mmaction-slowfast云端操作-纯享版

写在前面： 云端GPU使用AutoBL

一、步骤

1、清零文件夹

cd ../root/autodl-tmp/home/Custom-ava-dataset_Custom-Spatio-Temporally-Action-Video-Dataset/Dataset
rm -rf choose_frames choose_frames_all choose_frames_middle frames rawframes 
mkdir choose_frames choose_frames_all choose_frames_middle frames rawframes 

2、帧操作

填充choose_frames和choose_frames_all文件夹

bash cut_frames.sh 
python choose_frames_all.py 600 0
python choose_frames.py 600 0

600是视频长度，这里我的每个视频是10分钟。

3、YOLO操作

cd ../yolovDeepsort
rm -rf ./yolov5/runs
python ./yolov5/detect.py --source ../Dataset/choose_frames_all/ --save-txt --save-conf 
cd ./mywork
python dense_proposals_train.py ../yolov5/runs/detect/exp/labels ./dense_proposals_train.pkl show

YOLO用于产生dense_proposals_train.pkl文件

4、解压choose_frames_middle中的文件夹

cd ../../Dataset/choose_frames_middle
unzip choose_frames_middle.zip

5、提取标注完成的json文件

cd ../
python json_extract.py 

这一步执行后，生成train_without_personId.csv文件

6、通过deepsort检测图片

cd ../yolovDeepsort/mywork
python dense_proposals_train_deepsort.py ../yolov5/runs/detect/exp/labels ./dense_proposals_train_deepsort.pkl show
cd ../
python yolov5_to_deepsort.py --source ../Dataset/frames

7、融合actions与personReID

cd ../Dataset
python train_temp.py
python train.py

8、生成其他标注文件

cd ./annotations
touch train_excluded_timestamps.csv
cp ../../yolovDeepsort/mywork/dense_proposals_train.pkl .
cp ./dense_proposals_train.pkl dense_proposals_val.pkl
cp ./train.csv val.csv
cp train_excluded_timestamps.csv val_excluded_timestamps.csv
cd ../
cp -r ./frames/* rawframes
cd ../yolovDeepsort/mywork/
python change_raw_frames.py

9、修正文件

python change_dense_proposals_train.py
python change_dense_proposals_val.py

10、开始训练

cd ../../../mmaction2_YF
python tools/train.py configs/detection/ava/my_slowfast_kinetics_pretrained_r50_4x16x1_20e_ava_rgb.py --validate