Object Detection

1. Introduction

Adlik object_detection mainly studies the real-time and high-precision object detection algorithms,
and performs model compression and acceleration, so that they can be deployed on different
platforms and devices and achieve real-time performance.

At present, we mainly implement the training and model compression of the YOLOv4。 In the future, we will continue to explore real-time yolov4 methods that can maintain high accuracy on different hardware devices.

2. Update

2020.03.24

• YOLOv4 training that supports single-node or multi-node multi-GPU and mixed precision training.
• Pruning that supports regular pruning.
• Knowledge Distillation.

3. Enviroment

We recommend using Docker and have provided dockerfile files.

If you need to install it yourself, you need pytorch>=1.6 , python >=3.6 that uses CUDA 10.1.

• Create a python virtual environment and install the dependent packages.

pip install -r requirements.txt

• Clone the Adlik object detection code repository from github.

git clone https://github.com/Adlik/Adlik_object_detection.git

4. Training

Run commands below to reproduce results on COCO dataset or custom dataset.

• Single node multi-gpu (e.g. 8 gpu)

python -m torch.distributed.launch --nproc_per_node=8 train.py --cfg cfg/yolov4.cfg --data data/coco2017.yaml 
--weights weights/yolov4.conv.137 --epochs 200 --batch_size 4 --multi-scale

• multi node multi-gpu (e.g. 2 nodes and 8 gpu)

python -m torch.distributed.launch --nproc_per_node=8 --nnodes=2 --node_rank=0 --master_addr="master node ip" 
--master_port=11122 train.py --cfg cfg/yolov4.cfg --data data/coco2017.yaml --weights weights/yolov4.conv.137 
--epochs 200 --batch_size 6 --multi-scale

python -m torch.distributed.launch --nproc_per_node=8 --nnodes=2 --node_rank=1 --master_addr="master node ip" 
--master_port=11122 train.py --cfg cfg/yolov4.cfg --data data/coco2017.yaml --weights weights/yolov4.conv.137 
--epochs 200 --batch_size 6 --multi-scale

5. Model Compression

We will explore different model compression methods to enable real-time inference of models on different hardware platforms.

5.1 Pruning

Regular pruning adopts the pruning method of network slimming, which is designed for hardware deployment. The number of filters after pruning is a multiple of 8.

• Sparse training

-st means start sparse training.
-pf specifies the value of the penalty factor.

python -m torch.distributed.launch --nproc_per_node=8 train.py --cfg cfg/yolov4.cfg 
--data data/coco2017.yaml --weights weights/yolov4_last.pt --epochs 100 --batch_size 6 
-st -pf 0.0001

• Regular pruning

--percentspecifies the percentage of the number of pruned channels.

python regular_prune.py --cfg cfg/yolov4.cfg --data data/coco2017.yaml 
--weights weights/sparsity_last.pt --percent 0.3 --img_size 608

• Fine-tuning

python -m torch.distributed.launch --nproc_per_node=8 train.py --cfg cfg/prune_0.3_yolov4.cfg 
--data data/coco2017.yaml --weights weights/prune_0.3_sparsity_last.weights 
--batch_size 8 --epochs 100

5.2 Knowledge Distillation

If the precision of the model after fine-tuning pruning decreases more, we can take the
initial training network as the teacher network, and distill the knowledge of the pruned
network as the student network, so as to improve the precision of the model after pruning.

python -m torch.distributed.launch --nproc_per_node=8 train.py --cfg cfg/prune_0.3_yolov4.cfg 
--teacher_cfg cfg/yolov4.cfg --data data/coco2017.yaml --weights weights/prune_0.3_sparsity_last.weights 
--teacher_weights weights/yolov4_last.pt --batch_size 8 --epochs 100

6. Test results

6.1 COCO dataset

All the results are tested on val2017 dataset.

Model	size	[email protected]	[email protected]:0.95	GFLOPs	params	FPS(V100)
YOLOv4	608	0.664	0.427	128.7	246 M	53
Sparsity YOLOv4	608	0.647	0.419	128.7	246 M	53
Regular prune (30%)	608	0.015	-	92.8	136M	63
Knowledge distillation	608	0.614	0.360	92.8	136M	63