Tools Darknet AlexeyAB

[junxnone]

Reference

• Github repo
• How to train (to detect your custom objects)
• Datasets2Darknet
• paper - CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features
• paper - mixup - Bag of Freebies for Training Object Detection Neural Networks
• cfg files parameters
• Converting Yolo v3 models to TensorFlow and OpenVINO(IR) models

Brief

git clone https://github.com/AlexeyAB/darknet
cd darknet
make

Test

wget https://pjreddie.com/media/files/yolov3.weights
./darknet detector test ./cfg/coco.data ./cfg/yolov3.cfg ./yolov3.weights

Total BFLOPS 65.864
Loading weights from ./yolov3.weights...
 seen 64
Done!
Enter Image Path: data/horses.jpg
data/horses.jpg: Predicted in 12600.076000 milli-seconds.
horse: 88%
horse: 99%
horse: 96%
horse: 99%
Not compiled with OpenCV, saving to predictions.png instead

prediction

Build Config

可以编译 darknet 前通过修改 Makefile 添加如下支持

• NVIDIA
  • CUDA
  • CUDNN
• OPENCV
• OPENMP

-GPU=0
-CUDNN=0
+GPU=1
+CUDNN=1
-OPENMP=0
+OPENMP=1

Train with Pascal VOC

Prepare Config

• voc/train.txt - training images list
• voc/test.txt - test images list

wget http://pjreddie.com/media/files/darknet53.conv.74 
cd build/darknet/x64/data/voc/
wget http://pjreddie.com/media/files/VOCtrainval_11-May-2012.tar
wget http://pjreddie.com/media/files/VOCtrainval_06-Nov-2007.tar
wget http://pjreddie.com/media/files/VOCtest_06-Nov-2007.tar
tar vxf VOCtrainval_11-May-2012.tar
tar vxf VOCtrainval_06-Nov-2007.tar
tar vxf VOCtest_06-Nov-2007.tar
python3 voc_label.py
cat 2007_train.txt 2007_val.txt 2012_*.txt>train.txt

• cfg/yolov3-voc.cfg - YOLO Network arch and Training config

- batch=1
- subdivisions=1
+ batch=64
+ subdivisions=64

1 for Testing, 64 for Training

mixup=1
cutmix=1
mosaic=1
blur=1

• cfg/voc.data - your train/test images/labels and weight saved path
• cfg/voc.names - your classes

-train  = /home/pjreddie/data/voc/train.txt
-valid  = /home/pjreddie/data/voc/2007_test.txt
+train  = ./build/darknet/x64/data/voc/train.txt
+valid  = ./build/darknet/x64/data/voc/2007_test.txt
 names = data/voc.names
-backup = /home/pjreddie/backup/
+backup = ./backup/

• anchor cluster

darknet detector calc_anchors data/your.data -num_of_clusters 9 -width 416 -height 416

Training cmd

• Start Train ./darknet detector train cfg/voc.data cfg/yolov3-voc.cfg darknet53.conv.74
• Train with multi-GPU -gpus 0,1,2,3
• with map -map
• with Loss-chart -dont_show -mjpeg_port 8090

  loss-chart 需要编译 Makefile 中的 OPENCV 选项

./darknet detector train cfg/voc.data cfg/yolov3-voc.cfg darknet53.conv.74 -gpus 0,1,2,3 -map \
-dont_show -mjpeg_port 8090 

Stop and restart the train with saved weights

./darknet detector train cfg/voc.data cfg/yolov3-voc.cfg  \
backup/yolov3-voc_1008.weights \
-gpus 0,1,2,3 -map \
-dont_show -mjpeg_port 8090 

YOLOv3 mAP@0.5IOU@416

Test

• Test image

  ./darknet detector test ./data/your.data  ./your.cfg ./your.weights ./your_image

• Test with confidence threshold -thresh <val>

• Calculation mAP

  ./darknet detector map ./data/your.data  ./your.cfg ./your.weights 

• Calculation mAP with IoU threshold -iou_thresh <val> - [0.75 , 0.5 , 0.1]

OpenVINO

• https://github.com/mystic123/tensorflow-yolo-v3.git

• --tiny 指定转换 yolov3-tiny model

• python3 convert_weights_pb.py --class_names ../data/kaiquan.names \
  --data_format NHWC \
  --weights_file ../mixup_weights/yolov3-kaiquan_best.weights \
  python3 /opt/intel/openvino/deployment_tools/model_optimizer/mo_tf.py -b 1 \
  --input_model ./frozen_darknet_yolov3_model.pb \
  --tensorflow_use_custom_operations_config ./yolo_v3.json \
  --data_type FP32\
  --output_dir ./FP32
  python3 /opt/intel/openvino/deployment_tools/model_optimizer/mo_tf.py -b 1 \
  --input_model ./frozen_darknet_yolov3_model.pb \
  --tensorflow_use_custom_operations_config ./yolo_v3.json \
  --data_type FP16\
  --output_dir ./FP16

• benchmark

  
  source /opt/intel/openvino/bin/setupvars.sh
  python3 /opt/intel/openvino/deployment_tools/tools/benchmark_tool/benchmark_app.py \
   -i ../your_data/\
  -m FP16/frozen_darknet_yolov3_model.xml \
  -d MULTI:CPU,GPU,HDDL\
  -l /opt/intel/openvino/inference_engine/lib/intel64/libcpu_extension_avx2.so

# Tips
- YOLOv3 和 YOLOv3-SPP 都是用一样的 [pre-trained weights - darknet53.conv.74](https://pjreddie.com/media/files/darknet53.conv.74)
- YOLO labels format `class x y width height` - 相对于 `width/height` 的值 (0.0, 1.0]
  - 转换方法 - [voc_label.py](https://github.com/AlexeyAB/darknet/blob/master/scripts/voc_label.py#L12)

[junxnone]

junxnone/tech-io#54