Contents

Requirements

Features

Preparation

Training

Evaluation

Benchmarking the speed of network

---

Requirements

• python 3.6
• pytorch 0.4.0

Features

• Include both Yolov2 and Yolov3

• Good performance

  544x544	VOC2007 Test(mAP)	Time per forward (batch size = 1)
  Yolov2	77.6%	11.5ms
  Yolov3	79.6%	23.1ms

  The models are trained from pretrained weights on imagenet with this implementation.

• Train as fast as darknet

• A lot of efficient backbones on hand

  Like tiny yolov2, tiny yolov3, mobilenet, mobilenetv2, shufflenet(g2), shufflenetv2(1x), squeezenext(1.0-SqNxt-23v5), light xception, xception etc.

  Check folder vedanet/network/backbone for details.

  416x416	VOC2007 Test(mAP)	Time per forward (batch size = 1)
  TinyYolov2	57.5%	2.4ms
  TinyYolov3	61.3%	2.3ms

  The models are trained from scratch with this implementation.

---

Preparation

1) Code

git clone https://github.com/TencentYoutuResearch/ObjectDetection-OneStageDet

cd ObjectDetection-OneStageDet/yolo

yolo_root=$(pwd)

cd ${yolo_root}/utils/test

make -j32

2) Data

wget https://pjreddie.com/media/files/VOCtrainval_11-May-2012.tar

wget https://pjreddie.com/media/files/VOCtrainval_06-Nov-2007.tar

wget https://pjreddie.com/media/files/VOCtest_06-Nov-2007.tar

tar xf VOCtrainval_11-May-2012.tar

tar xf VOCtrainval_06-Nov-2007.tar

tar xf VOCtest_06-Nov-2007.tar

cd VOCdevkit

VOCdevkit_root=$(pwd)

There will now be a VOCdevkit subdirectory with all the VOC training data in it.

mkdir ${VOCdevkit_root}/onedet_cache

cd ${yolo_root}

open examples/labels.py, let the variable ROOT point to ${VOCdevkit_root}

python examples/labels.py

open cfgs/yolov2.yml, let the data_root_dir point to ${VOCdevkit_root}/onedet_cache

open cfgs/yolov3.yml, let the data_root_dir point to ${VOCdevkit_root}/onedet_cache

3) weights

Download model weights from baidudrive or googledrive.

Or downlowd darknet19_448.conv.23 and darknet53.conv.74 from darknet website:

wget https://pjreddie.com/media/files/darknet19_448.conv.23

wget https://pjreddie.com/media/files/darknet53.conv.74

Then, move all the model weights to ${yolo_root}/weights directory.

---

Training

cd ${yolo_root}

1) Yolov2

1.1) open cfgs/yolov2.yml, let the weights of train block point to the pretrain weights

1.2) open cfgs/yolov2.yml, let the gpus of train block point to an available gpu id

1.3) If you want to print log onto screen, make the stdout of train block True in cfgs/yolov2.yml

1.4) run

python examples/train.py Yolov2

2) Yolov3

2.1) open cfgs/yolov3.yml, let the weights of train block point to the pretrain weights

2.2) open cfgs/yolov3.yml, let the gpus of train block point to an available gpu id

2.3) If you want to print log onto screen, make the stdout of train block True in cfgs/yolov3.yml

2.4) run

python examples/train.py Yolov3

3) Results

The logs and weights will be in ${yolo_root}/outputs.

4) Other models

There are many other models like tiny yolov2, tiny yolov3, mobilenet, mobilenetv2, shufflenet(g2), shufflenetv2(1x), squeezenext(1.0-SqNxt-23v5), light xception, xception etc. You can try these like 1) Yolov2 part.

---

Evaluation

cd ${yolo_root}

1) Yolov2

1.1) open cfgs/yolov2.yml, let the gpus of test block point to an available gpu id

1.2) run

python examples/test.py Yolov2

2) Yolov3

2.1) open cfgs/yolov3.yml, let the gpus of test block point to an available gpu id

2.2) run

python examples/test.py Yolov3

3) Results

The output bbox will be in ${yolo_root}/results, every line of the file in ${yolo_root}/results has a format like img_name confidence xmin ymin xmax ymax

4) Other models

There are many other models like tiny yolov2, tiny yolov3, mobilenet, mobilenetv2, shufflenet(g2), shufflenetv2(1x), squeezenext(1.0-SqNxt-23v5), light xception, xception etc. You can try these like 1) Yolov2 part.

---

Benchmarking the speed of network

cd ${yolo_root}

1) Yolov2

1.1) open cfgs/yolov2.yml, let the gpus of speed block point to an available gpu id

1.2) run

python examples/speed.py Yolov2

2) Yolov3

2.1) open cfgs/yolov3.yml, let the gpus of speed block point to an available gpu id

2.2) run

python examples/speed.py Yolov3

3) Tiny Yolov2

3.1) open cfgs/tiny_yolov2.yml, let the gpus of speed block point to an available gpu id

3.2) run

python examples/speed.py TinyYolov2

4) Tiny Yolov3

4.1) open cfgs/tiny_yolov3.yml, let the gpus of speed block point to an available gpu id

4.2) run

python examples/speed.py TinyYolov3

5) Mobilenet

5.1) open cfgs/region_mobilenet.yml, let the gpus of speed block point to an available gpu id

5.2) run

python examples/speed.py RegionMobilenet

6) Other backbones with region loss

You can try these like 5) Mobilenet part.

---

Credits

I got a lot of code from lightnet, thanks to EAVISE.