gulp>> docker-darknet-yolov3-yolov4-training-test>> 返回
项目作者： hanifalisohag

项目描述 ：
Docker Darknet YOLOv3 YOLOv4 Training and Test (With GPU CUDA support on Linux)
高级语言： Dockerfile
项目地址: git://github.com/hanifalisohag/docker-darknet-yolov3-yolov4-training-test.git


Docker Darknet YOLOv3 YOLOv4 Training and Test (With GPU CUDA support on Linux)

Build it

  1. $ docker build -t docker-darknet_yolo:latest .

Run it

without X-server

  1. $ docker run -it --gpus all docker-darknet_yolo:latest

with X-server

  1. $ docker run -it --gpus all --net host -e DISPLAY=$DISPLAY -v $HOME/.Xauthority:/root/.Xauthority docker-darknet_yolo:latest

Test it

YOLOv4

  1. $ ./darknet detector test ./cfg/coco.data ./cfg/yolov4.cfg yolov4.weights data/dog.jpg -dont_show

YOLOv3

  1. $ ./darknet detector test ./cfg/coco.data ./cfg/yolov3.cfg yolov3.weights data/dog.jpg -dont_show

For training:

On the host machine

  1. Clone the repository 
    1. $ git clone https://github.com/hanifalisohag/docker-darknet-yolov3-yolov4-training-test.git

  2. Put all your custom “.jpg” and “.txt” annotations on this folder : /dataset/data/obj

  3. Generate the val.txt and train.txt on the /dataset/data folder 

    1. $ cd dataset
    2. $ python train_test_split.py
    3. OR
    4. $ python3 train_test_split.py
  4. Please change the yolo-obj.cfg, obj.data and yolo-obj.names file according to your custom class numbers and YOLOv3 and YOLOv4. This repo has been configured of 6 classes for YOLOv3
  5. The full dataset direcotry structures should look like below-
    1. ├── data
    2. │   ├── obj
    3. │   │   ├── 000000000036.jpg
    4. │   │   ├── 000000000036.txt
    5. │   │   ├── 000000581921.jpg
    6. │   │   ├── 000000581921.txt
    7. │   │   └── ................
    8. │   ├── train.txt
    9. │   ├── val.txt
    10. │   └── yolo-obj.names
    11. ├── obj.data
    12. ├── train_test_split.py
    13. └── yolo-obj.cfg

    Run the following command to start the training (Manually)

    ```
    $ docker run -it —gpus all —name darknet_training\
    -v ‘$(PWD)/dataset/data/obj:/darknet/data/obj’ \
    -v ‘$(PWD)/dataset/data/val.txt:/darknet/data/val.txt’ \
    -v ‘$(PWD)/dataset/data/train.txt:/darknet/data/train.txt’ \
    -v ‘$(PWD)/dataset/data/yolo-obj.names:/darknet/data/yolo-obj.names’ \
    -v ‘$(PWD)/dataset/obj.data:/darknet/obj.data’ \
    -v ‘$(PWD)/dataset/yolo-obj.cfg:/darknet/yolo-obj.cfg’ \
    -v ‘$(PWD)/dataset/backup:/darknet/backup’ \
    docker-darknet_yolo:latest /bin/bash
  1. ### Inside the running container (/darknet#):
  2. #### YOLOv3

./darknet detector train obj.data yolo-obj.cfg darknet53.conv.74 -map -dont_show

  1. #### YOLOv4

./darknet detector train obj.data yolo-obj.cfg yolov4.conv.137 -map -dont_show

  1. ## Running the container in detached mode:
  2. #### YOLOv3

$ docker run -it —gpus all —name darknet_training -d\
-v ‘$(PWD)/dataset/data/obj:/darknet/data/obj’ \
-v ‘$(PWD)/dataset/data/val.txt:/darknet/data/val.txt’ \
-v ‘$(PWD)/dataset/data/train.txt:/darknet/data/train.txt’ \
-v ‘$(PWD)/dataset/data/yolo-obj.names:/darknet/data/yolo-obj.names’ \
-v ‘$(PWD)/dataset/obj.data:/darknet/obj.data’ \
-v ‘$(PWD)/dataset/yolo-obj.cfg:/darknet/yolo-obj.cfg’ \
-v ‘$(PWD)/dataset/backup:/darknet/backup’ \
docker-darknet_yolo:latest ./darknet detector train obj.data yolo-obj.cfg darknet53.conv.74 -map -dont_show

  2. #### YOLOv4

$ docker run -it —gpus all —name darknet_training -d\
-v ‘$(PWD)/dataset/data/obj:/darknet/data/obj’ \
-v ‘$(PWD)/dataset/data/val.txt:/darknet/data/val.txt’ \
-v ‘$(PWD)/dataset/data/train.txt:/darknet/data/train.txt’ \
-v ‘$(PWD)/dataset/data/yolo-obj.names:/darknet/data/yolo-obj.names’ \
-v ‘$(PWD)/dataset/obj.data:/darknet/obj.data’ \
-v ‘$(PWD)/dataset/yolo-obj.cfg:/darknet/yolo-obj.cfg’ \
-v ‘$(PWD)/dataset/backup:/darknet/backup’ \
docker-darknet_yolo:latest ./darknet detector train obj.data yolo-obj.cfg yolov4.conv.137 -map -dont_show

  2. ### To attach to the container

$docker attach —detach-keys=”ctrl-a,x” darknet_training

  1. In order to get out of the container without exiting it **press** `CTRL+A` then **press** `X`. It will get the user out of the running container.
  5. ### Starting a stopped container
  6. First get the docker container ID/Name

$ docker ps -a

  1. It will provide the ID/Name of the docker container we are interested. In this example **darknet_training**
  2. Now you can start the stopped ***darknet_training*** container using the follwing commands-

$ docker start darknet_training
$ docker exec -it darknet_training /bin/sh

  2. ### To delete that container

$ docker rm -f darknet_training

  2. ## To continue a finished training
  3. We can change the iteration number of the config file located at **/dataset/yolo-obj.cfg** and restart the training from where we stopped or finised training (Suppose, it was **100200**). You need to change the following lines 150200 and steps would be 80% (120160), 90% (135180)

max_batches = 150200
policy=steps
steps=120160,135180

  2. After that, remove the stopped container named **darknet_training**

$ docker rm -f darknet_training

  2. Now, to start the training run the following command again. Please change the last iteration weights file accordingly. In this example, **backup/yolo-obj_100000.weights**, since the previous iterations were **100200** and the previous training saved the **yolo-obj_100000.weights** in the backup folder.

$ docker run -it —gpus all —name darknet_training -d\
-v ‘$(PWD)/dataset/data/obj:/darknet/data/obj’ \
-v ‘$(PWD)/dataset/data/val.txt:/darknet/data/val.txt’ \
-v ‘$(PWD)/dataset/data/train.txt:/darknet/data/train.txt’ \
-v ‘$(PWD)/dataset/data/yolo-obj.names:/darknet/data/yolo-obj.names’ \
-v ‘$(PWD)/dataset/obj.data:/darknet/obj.data’ \
-v ‘$(PWD)dataset/yolo-obj.cfg:/darknet/yolo-obj.cfg’ \
-v ‘$(PWD)/dataset/backup:/darknet/backup’ \
docker-darknet_yolo:latest ./darknet detector train obj.data yolo-obj.cfg backup/yolo-obj_100000.weights -map

  2. ## Building on a Xavier

$ sudo docker build -f Dockerfile_jetson -t docker-darknet_yolo:latest .

$ sudo docker run -it —runtime=nvidia docker-darknet_yolo:latest

run inside the docker container to succefully build darknet with NVIDIA GPU Runtime

$ root@4b397cef37d7:/darknet# make
$ exit

To get the container ID (then check for the last stopped/exited container)

$ sudo docker ps -a

$ sudo docker commit 4b397cef37d7 docker-darknet_yolo:latest

  2. ## Running on a Xavier

$ sudo docker run -it —runtime=nvidia docker-darknet_yolo:latest

inside the container

$ ./darknet detector test ./cfg/coco.data ./cfg/yolov4.cfg yolov4.weights data/dog.jpg -dont_show

  2. ## Running on a Xavier with GUI

$ xhost +

With GUI X11

$ sudo docker run -it —runtime=nvidia —net host -e DISPLAY=$DISPLAY -v /tmp/.X11-unix/:/tmp/.X11-unix docker-darknet_yolo:latest

inside the container

$ ./darknet detector test ./cfg/coco.data ./cfg/yolov4.cfg yolov4.weights data/dog.jpg

```

Output

Alt text