YOLOv4: Darknet 如何于 Docker 编译，及训练 COCO 子集

YOLO 算法是非常著名的目标检测算法。从其全称 You Only Look Once: Unified, Real-Time Object Detection ，可以看出它的特性：

Look Once: one-stage (one-shot object detectors) 算法，把目标检测的两个任务分类和定位一步完成。
Unified: 统一的架构，提供 end-to-end 的训练和预测。
Real-Time: 实时性，初代论文给出的指标 FPS 45 , mAP 63.4 。

YOLOv4: Optimal Speed and Accuracy of Object Detection ，于今年 4 月公布，采用了很多近些年 CNN 领域优秀的优化技巧。其平衡了精度与速度，目前在实时目标检测算法中精度是最高的。

论文地址：

YOLO: https://arxiv.org/abs/1506.02640
YOLO v4: https://arxiv.org/abs/2004.10934

源码地址：

YOLO: https://github.com/pjreddie/d...
YOLO v4: https://github.com/AlexeyAB/d...

本文将介绍 YOLOv4 官方 Darknet 实现，如何于 Docker 编译使用。以及从 MS COCO 2017 数据集中怎么选出部分物体，训练出模型。

主要内容有：

准备 Docker 镜像
准备 COCO 数据集
用预训练模型进行推断
准备 COCO 数据子集
训练自己的模型并推断
参考内容

准备 Docker 镜像

首先，准备 Docker ，请见：Docker: Nvidia Driver, Nvidia Docker 推荐安装步骤。

之后，开始准备镜像，从下到上的层级为：

nvidia/cuda: https://hub.docker.com/r/nvid...
OpenCV: https://github.com/opencv/opencv
Darknet: https://github.com/AlexeyAB/d...

nvidia/cuda

准备 Nvidia 基础 CUDA 镜像。这里我们选择 CUDA 10.2 ，不用最新 CUDA 11，因为现在 PyTorch 等都还都是 10.2 呢。

拉取镜像：

docker pull nvidia/cuda:10.2-cudnn7-devel-ubuntu18.04

测试镜像：

$ docker run --gpus all nvidia/cuda:10.2-cudnn7-devel-ubuntu18.04 nvidia-smi
Sun Aug 8 00:00:00 2020
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 440.100 Driver Version: 440.100 CUDA Version: 10.2 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
|===============================+======================+======================|
| 0 GeForce RTX 208... Off | 00000000:07:00.0 On | N/A |
| 0% 48C P8 14W / 300W | 340MiB / 11016MiB | 2% Default |
+-------------------------------+----------------------+----------------------+
| 1 GeForce RTX 208... Off | 00000000:08:00.0 Off | N/A |
| 0% 45C P8 19W / 300W | 1MiB / 11019MiB | 0% Default |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: GPU Memory |
| GPU PID Type Process name Usage |
|=============================================================================|
+-----------------------------------------------------------------------------+

OpenCV

基于 nvidia/cuda 镜像，构建 OpenCV 的镜像：

cd docker/ubuntu18.04-cuda10.2/opencv4.4.0/

docker build 
-t joinaero/ubuntu18.04-cuda10.2:opencv4.4.0 
--build-arg opencv_ver=4.4.0 
--build-arg opencv_url=https://gitee.com/cubone/opencv.git 
--build-arg opencv_contrib_url=https://gitee.com/cubone/opencv_contrib.git 
.

其 Dockerfile 可见这里： https://github.com/ikuokuo/st... 。

Darknet

基于 OpenCV 镜像，构建 Darknet 镜像：

cd docker/ubuntu18.04-cuda10.2/opencv4.4.0/darknet/

docker build 
-t joinaero/ubuntu18.04-cuda10.2:opencv4.4.0-darknet 
.

其 Dockerfile 可见这里： https://github.com/ikuokuo/st... 。

上述镜像已上传 Docker Hub 。如果 Nvidia 驱动能够支持 CUDA 10.2 ，那可以直接拉取该镜像：

docker pull joinaero/ubuntu18.04-cuda10.2:opencv4.4.0-darknet

准备 COCO 数据集

MS COCO 2017 下载地址: http://cocodataset.org/#download

图像，包括：

2017 Train images [118K/18GB]
- http://images.cocodataset.org...
2017 Val images [5K/1GB]
- http://images.cocodataset.org...
2017 Test images [41K/6GB]
- http://images.cocodataset.org...
2017 Unlabeled images [123K/19GB]
- http://images.cocodataset.org...

标注，包括：

2017 Train/Val annotations [241MB]
- http://images.cocodataset.org...
2017 Stuff Train/Val annotations [1.1GB]
- http://images.cocodataset.org...
2017 Panoptic Train/Val annotations [821MB]
- http://images.cocodataset.org...
2017 Testing Image info [1MB]
- http://images.cocodataset.org...
2017 Unlabeled Image info [4MB]
- http://images.cocodataset.org...

用预训练模型进行推断

预训练模型 yolov4.weights ，下载地址 https://github.com/AlexeyAB/d... 。

运行镜像：

xhost +local:docker

docker run -it --gpus all 
-e DISPLAY 
-e QT_X11_NO_MITSHM=1 
-v /tmp/.X11-unix:/tmp/.X11-unix 
-v $HOME/.Xauthority:/root/.Xauthority 
--name darknet 
--mount type=bind,source=$HOME/Codes/devel/datasets/coco2017,target=/home/coco2017 
--mount type=bind,source=$HOME/Codes/devel/models/yolov4,target=/home/yolov4 
joinaero/ubuntu18.04-cuda10.2:opencv4.4.0-darknet

进行推断：

./darknet detector test cfg/coco.data cfg/yolov4.cfg /home/yolov4/yolov4.weights 
-thresh 0.25 -ext_output -show -out /home/coco2017/result.json 
/home/coco2017/test2017/000000000001.jpg

推断结果：

准备 COCO 数据子集

MS COCO 2017 数据集有 80 个物体标签。我们从中选取自己关注的物体，重组个子数据集。

首先，获取样例代码：

git clone https://github.com/ikuokuo/start-yolov4.git

scripts/coco2yolo.py: COCO 数据集转 YOLO 数据集的脚本
scripts/coco/label.py: COCO 数据集的物体标签有哪些
cfg/coco/coco.names: 编辑我们想要的那些物体标签

之后，准备数据集：

cd start-yolov4/
pip install -r scripts/requirements.txt

export COCO_DIR=$HOME/Codes/devel/datasets/coco2017

# train
python scripts/coco2yolo.py 
--coco_img_dir $COCO_DIR/train2017/ 
--coco_ann_file $COCO_DIR/annotations/instances_train2017.json 
--yolo_names_file ./cfg/coco/coco.names 
--output_dir ~/yolov4/coco2017/ 
--output_name train2017 
--output_img_prefix /home/yolov4/coco2017/train2017/

# valid
python scripts/coco2yolo.py 
--coco_img_dir $COCO_DIR/val2017/ 
--coco_ann_file $COCO_DIR/annotations/instances_val2017.json 
--yolo_names_file ./cfg/coco/coco.names 
--output_dir ~/yolov4/coco2017/ 
--output_name val2017 
--output_img_prefix /home/yolov4/coco2017/val2017/

数据集，内容如下：

~/yolov4/coco2017/
├── train2017/
│ ├── 000000000071.jpg
│ ├── 000000000071.txt
│ ├── ...
│ ├── 000000581899.jpg
│ └── 000000581899.txt
├── train2017.txt
├── val2017/
│ ├── 000000001353.jpg
│ ├── 000000001353.txt
│ ├── ...
│ ├── 000000579818.jpg
│ └── 000000579818.txt
└── val2017.txt

训练自己的模型并推断

准备必要文件

cfg/coco/coco.names <cfg/coco/coco.names.bak has original 80 objects>
- Edit: keep desired objects
cfg/coco/yolov4.cfg <cfg/coco/yolov4.cfg.bak is original file>
- Download yolov4.cfg, then changed:
- batch=64, subdivisions=32 <32 for 8-12 GB GPU-VRAM>
- width=512, height=512
- classes=
- max_batches=
- steps=0.8, max_batches0.9>
- filters=<(classes+5)x3, in the 3 [convolutional] before each [yolo] layer>
- ~~filters=<(classes+9)x3, in the 3 [convolutional] before each [Gaussian_yolo] layer>~~
cfg/coco/coco.data
- Edit: train, valid to YOLO datas
csdarknet53-omega.conv.105
- Download csdarknet53-omega_final.weights, then run:

docker run -it --rm --gpus all 
--mount type=bind,source=$HOME/Codes/devel/models/yolov4,target=/home/yolov4 
joinaero/ubuntu18.04-cuda10.2:opencv4.4.0-darknet

./darknet partial cfg/csdarknet53-omega.cfg /home/yolov4/csdarknet53-omega_final.weights /home/yolov4/csdarknet53-omega.conv.105 105

训练自己的模型

运行镜像：

cd start-yolov4/

xhost +local:docker

docker run -it --gpus all 
-e DISPLAY 
-e QT_X11_NO_MITSHM=1 
-v /tmp/.X11-unix:/tmp/.X11-unix 
-v $HOME/.Xauthority:/root/.Xauthority 
--name darknet 
--mount type=bind,source=$HOME/Codes/devel/models/yolov4,target=/home/yolov4 
--mount type=bind,source=$HOME/yolov4/coco2017,target=/home/yolov4/coco2017 
--mount type=bind,source=$PWD/cfg/coco,target=/home/cfg 
joinaero/ubuntu18.04-cuda10.2:opencv4.4.0-darknet

进行训练：

mkdir -p /home/yolov4/coco2017/backup

# Training command
./darknet detector train /home/cfg/coco.data /home/cfg/yolov4.cfg /home/yolov4/csdarknet53-omega.conv.105 -map

中途可以中断训练，然后这样继续：

# Continue training
./darknet detector train /home/cfg/coco.data /home/cfg/yolov4.cfg /home/yolov4/coco2017/backup/yolov4_last.weights -map

yolov4_last.weights 每迭代 100 次，会被记录。

如果多 GPU 训练，可以在 1000 次迭代后，加参数 -gpus 0,1 ，再继续：

# How to train with multi-GPU
# 1. Train it first on 1 GPU for like 1000 iterations
# 2. Then stop and by using partially-trained model `/backup/yolov4_1000.weights` run training with multigpu
./darknet detector train /home/cfg/coco.data /home/cfg/yolov4.cfg /home/yolov4/coco2017/backup/yolov4_1000.weights -gpus 0,1 -map

训练过程，记录如下：

加参数 -map 后，上图会显示有红线 mAP。

查看模型 mAP@IoU=50 精度：

$ ./darknet detector map /home/cfg/coco.data /home/cfg/yolov4.cfg /home/yolov4/coco2017/backup/yolov4_final.weights

...
Loading weights from /home/yolov4/coco2017/backup/yolov4_final.weights...
seen 64, trained: 384 K-images (6 Kilo-batches_64)
Done! Loaded 162 layers from weights-file
  
calculation mAP (mean average precision)...
Detection layer: 139 - type = 27
Detection layer: 150 - type = 27
Detection layer: 161 - type = 27
160
detections_count = 745, unique_truth_count = 190
class_id = 0, name = train, ap = 80.61% (TP = 142, FP = 18)

for conf_thresh = 0.25, precision = 0.89, recall = 0.75, F1-score = 0.81
for conf_thresh = 0.25, TP = 142, FP = 18, FN = 48, average IoU = 75.31 %

IoU threshold = 50 %, used Area-Under-Curve for each unique Recall

mean average precision ([email protected]) = 0.806070, or 80.61 %
Total Detection Time: 4 Seconds