Model_Compression_For_YOLOV4

In this repository using the dynamic sparse training( variable sparse rate s which can speed up the sparse training process), channel pruning and knowledge distilling for YOLOV3 and YOLOV4;

And if you like it, give it star.

1. YOLOv3v4 -ModelCompression-Training

This repository project mainly include three parts.

• Part1. Common training and sparse training(prepare for the channel pruning) for object detection datasets(COCO2017, VOC, OxforHand).

• Part2. General model compression algorithm including pruning and knowledge distillation.

• Part3. A brief introduce for Network quantization .

Source code using Pytorch implementation to ultralytics/yolov3 for yolov3 source code.

For the YOLOV4 pytorch version, try this /~https://github.com/Tianxiaomo/pytorch-YOLOv4.

2. Datasets and Environment Requirements

Make a COCO or VOC dataset for this project try here dataset_for_Ultralytics_training.

The environment is Pytorch >= 1.1.0 , see the ./requiremnts.txt and also can reference the ultralytics/yolov3 ./requirements.txt .

---

3. Part1. Common training and sparse training

here sparse training on object detection datasets is prepared for the channel pruning .

3.1 For the common training use the following command:

python3 train.py --data ... --cfg ... -pt --weights ... --img_size ... --batch-size ... --epochs ... : -pt means that will use the pretrained model's weight.

3.2 For the sparse training use the:

python3 train.py --data ... -sr --s 0.001 --prune 0  -pt --weights ... --cfg ... --img_size ...  --batch-size 32  --epochs ...

3.3 parameter explaination:

-sr: Sparse training,--s: Specifies the sparsity factor，--prune :Specify the sparsity type.

--prune 0 is the sparsity of normal pruning and regular pruning.

--prune 1 is the sparsity of shortcut pruning.

--prune 2 is the sparsity of layer pruning.

-details see the 2.1.

3.4 Notice for the sparse training:

-The reason for using sparse training before we prune the network is that we need to select out the unimportant channels in the network, through the sparse training we can select out and prune these unimportant channels in the network.

-When the classes you trian network is not too much, such 1-5 classes. There maybe no difference with sparse training or without sparse training before prune the network.

-When the training classes are above 10 clasees, sparse training play an important role, in this situation prune the channel directly without sparse training will bring an irreparable damage to the network's accuracy, even later use the fine-tune or distillation it brings a little effect. Meanwhile, doing the sparse training firstly, then prune the network it may reduce the network's accuracy temporary, after we fine-tune or distilling the pruned network, the pruned network's accuracy will be regained.

python3 train.py --data ... -sr --s 0.001 --prune 0  -pt --weights ... --cfg ... --img_size ...  --batch-size 32  --epochs ...

3.5 The original weights distribution and sparse training weights distribution

Fig1: left- the original weight distribution right: the weight distribution after sprse training

Fig2: The mAP during the sparse training on the VOC dataset

3.6 Testing and detect command:

python3 test.py --data ... --cfg ... : Test the mAP@0.5 command

python3 detect.py --data ... --cfg ... --source ...: Detection a single image/video command, default address of source is data/samples, the output result is saved in the output file.

---

4. Part2 Model compression algorithm

this part included pruning and knowledge distillation.

4.1 channel pruning types

method	advantage	disadvantage
Normal pruning	Not prune for shortcut layer. It has a considerable and stable compression rate that requires no fine tuning.	The compression rate is not extreme.
Shortcut pruning	Very high compression rate.	Fine-tuning is necessary.
Silmming	Shortcut fusion method was used to improve the precision of shear planting.	Best way for shortcut pruning
Regular pruning	Designed for hardware deployment, the number of filters after pruning is a multiple of 2, no fine-tuning, support tiny-yolov3 and Mobilenet series.	Part of the compression ratio is sacrificed for regularization.
layer pruning	ResBlock is used as the basic unit for purning, which is conducive to hardware deployment.	It can only cut backbone.
layer-channel pruning	First, use channel pruning and then use layer pruning, and pruning rate was very high.	Accuracy may be affected.

4.2 Pruning the network command:

-for the channel pruning types:

python3 normal(or regular/shortcut/slim)_prune.py --cfg ... --data ... --weights ... --percent ...

-for the layer pruning(it is actually based on the channel pruning):

python3 layer_prune.py --cfg ... --data ... --weights ... --shortcut ...

python3 layer_channel_prune.py --cfg ... --data ... --weights ... --shortcut ... --percent ...

-Notice that we can get more compression by increasing the percent value, but if the sparsity is not enough and the percent value is too high, the program will report an error.

4.3 Network Knowledge distillation:

4.3.1 Important Notice

• For the pruned model, we can fine tune 20~ 50 epochs to recover the pruned model's accuracy!
• After that we use the pruned and fine-tuned model as student network, the original network(before sparse training ) to do the knowledge distillation.

-The basic distillation method Distilling the Knowledge in a Neural Network was proposed by Hinton in 2015, and has been partially improved in combination with the detection network.

4.3.2 Knowledge distillation command, add the --t_cfg --t_weights --KDstr choice:

python train.py --data ... --batch-size ... --weights ... --cfg ... --img-size ... --epochs ... --t_cfg ... --t_weights ...

--t_cfg :cfg file of teacher model --t_weights: weights file of teacher model --KDstr :KD strategy

`--KDstr 1` KLloss can be obtained directly from the output of teacher network and the output of student network and added to the overall loss.
`--KDstr 2` To distinguish between box loss and class loss, the student does not learn directly from the teacher. L2 distance is calculated respectively for student, teacher and GT. When student is greater than teacher, an additional loss is added for student and GT.
`--KDstr 3` To distinguish between Boxloss and ClassLoss, the student learns directly from the teacher.
`--KDstr 4` KDloss is divided into three categories, box loss, class loss and feature loss.
`--KDstr 5` On the basis of KDstr 4, the fine-grain-mask is added into the feature

Usually, the original(or unpruned model but has been sparse trained) model is used as the teacher model, and the post-compression model is used as the student model for distillation training to improve the mAP of student network.

---

5 Part3. A brief introduce for Network quantization

5.1 Due to the model weight has been quantized from FP32 to INT8:

• Most our personal PC machine can not run the quantized model with this int8 data type.

• And quantization method usually co-operate with specific hardware platform, such as Xilinx use Vitis Ai to quantize the model and deploy on the Zynq-ultraScale series(like pynq-z2, ultra_96_v2, ZCU104); Nvidia use TensorRT to quantize the model and deploy it on the Jetson (like Nano, TX1, TX2) ;

• Here are the reference On Ultra_96_v2, On Jetson Nano we use their tools quantize our pruned yolov4 network and deploy it on thier hardware target.

5.2 Recently, the Pytorch 1.8 has launch a "torch.fx" module:

-This would be a fortune for us to reasearch on the quantization;

-We wish we can bring the other repository that focus on quantization, God bless us.

5.3 Quantize command:

--quantized 2 Dorefa quantization method

python train.py --data ... --batch-size ... --weights ... --cfg ... --img-size ... --epochs ... --quantized 2

--quantized 1 Google quantization method

python train.py --data ... --batch-size ... --weights ... --cfg ... --img-size ... --epochs ... --quantized 3

--BN_Flod using BN Flod training, --FPGA Pow(2) quantization for FPGA.

Reference:

---

-Papers: Pruning method based on BN layer comes from Learning Efficient Convolutional Networks through Network Slimming.

Pruning without fine-tune Rethinking the Smaller-Norm-Less-Informative Assumption in Channel Pruning of Convolution Layers.

Attenton transfer distilling Paying More Attention to Attention: Improving the Performance of Convolutional Neural Networks via Attention Transfer

-Repositories: Channel pruning method based on BN layers for the yolov3 and yolov4, we recommond the following repository:

/~https://github.com/tanluren/yolov3-channel-and-layer-pruning

coldlarry/YOLOv3-complete-pruning

/~https://github.com/SpursLipu/YOLOv3v4-ModelCompression-MultidatasetTraining-Multibackbone

Thanks for your contributions.

Further reading

Here is our paper Group channel pruning and spatial attention distilling for object detection

@article{chu2022group,

title={Group channel pruning and spatial attention distilling for object detection},

author={Chu, Yun and Li, Pu and Bai, Yong and Hu, Zhuhua and Chen, Yongqing and Lu, Jiafeng},

journal={Applied Intelligence},

pages={1--19}, year={2022},

publisher={Springer} }