/DAMO-YOLO

DAMO-YOLO: a fast and accurate object detection method with some new techs, including NAS backbones, efficient RepGFPN, ZeroHead, AlignedOTA, and distillation enhancement.

Primary LanguagePythonApache License 2.0Apache-2.0

English | 简体中文

Contributing README-cn ThirdParty IndustryModels

Introduction

Welcome to DAMO-YOLO! It is a fast and accurate object detection method, which is developed by TinyML Team from Alibaba DAMO Data Analytics and Intelligence Lab. And it achieves a higher performance than state-of-the-art YOLO series. DAMO-YOLO is extend from YOLO but with some new techs, including Neural Architecture Search (NAS) backbones, efficient Reparameterized Generalized-FPN (RepGFPN), a lightweight head with AlignedOTA label assignment, and distillation enhancement. For more details, please refer to our Arxiv Report. Moreover, here you can find not only powerful models, but also highly efficient training strategies and complete tools from training to deployment.

Updates

  • [2023/04/12: We release DAMO-YOLO v0.3.1!] new
    • Add 701-categories DAMO-YOLO-S model, which cover more application scenarios and serve as high-quality pre-training model to improve the performance of downstream tasks。
    • Upgrade the DAMO-YOLO-Nano series model, which achieves 32.3/38.2/40.5 mAP with only 1.56/3.69/6.04 Flops, and runs in real-time at 4.08/5.05/6.69ms using Intel-CPU.
    • Add DAMO-YOLO-L model, which achieves 51.9 mAP with 7.95ms latency using T4-GPU.
  • [2023/03/13: We release DAMO-YOLO v0.3.0!]
    • Release DAMO-YOLO-Nano, which achieves 35.1 mAP with only 3.02GFlops.
    • Upgrade the optimizer builder, edits the optimizer config, you are able to use any optimizer supported by Pytorch.
    • Upgrade the data loading pipeline and training parameters, leading to significant improvements of DAMO-YOLO models, e.g., the mAP of DAMO-YOLO-T/S/M increased from 43.0/46.8/50.0 to 43.6/47.7/50.2 respectively.
  • [2023/02/15: Baseline for The 3rd Anti-UAV Challenge.]
  • [2023/01/07: We release DAMO-YOLO v0.2.1!]
  • [2022/12/15: We release DAMO-YOLO v0.1.1!]

Web Demo

Model Zoo

General Models

Model size mAPval
0.5:0.95
Latency T4
TRT-FP16-BS1
FLOPs
(G)
Params
(M)
AliYun Download Google Download
DAMO-YOLO-T 640 42.0 2.78 18.1 8.5 torch,onnx --
DAMO-YOLO-T* 640 43.6 2.78 18.1 8.5 torch,onnx --
DAMO-YOLO-S 640 46.0 3.83 37.8 16.3 torch,onnx --
DAMO-YOLO-S* 640 47.7 3.83 37.8 16.3 torch,onnx --
DAMO-YOLO-M 640 49.2 5.62 61.8 28.2 torch,onnx --
DAMO-YOLO-M* 640 50.2 5.62 61.8 28.2 torch,onnx --
DAMO-YOLO-L 640 50.8 7.95 97.3 42.1 torch,onnx --
DAMO-YOLO-L* 640 51.9 7.95 97.3 42.1 torch,onnx --
Legacy models
Model size mAPval
0.5:0.95
Latency T4
TRT-FP16-BS1
FLOPs
(G)
Params
(M)
AliYun Download Google Download
DAMO-YOLO-T 640 41.8 2.78 18.1 8.5 torch,onnx torch,onnx
DAMO-YOLO-T* 640 43.0 2.78 18.1 8.5 torch,onnx torch,onnx
DAMO-YOLO-S 640 45.6 3.83 37.8 16.3 torch,onnx torch,onnx
DAMO-YOLO-S* 640 46.8 3.83 37.8 16.3 torch,onnx torch,onnx
DAMO-YOLO-M 640 48.7 5.62 61.8 28.2 torch,onnx torch,onnx
DAMO-YOLO-M* 640 50.0 5.62 61.8 28.2 torch,onnx torch,onnx
  • We report the mAP of models on COCO2017 validation set, with multi-class NMS.
  • The latency in this table is measured without post-processing(NMS).
  • * denotes the model trained with distillation.
  • We use S as teacher to distill T, and M as teacher to distill S, ans L as teacher to distill M, while L is distilled by it self.

Light Models

Model size mAPval
0.5:0.95
Latency(ms) CPU
OpenVino-Intel8163
FLOPs
(G)
Params
(M)
AliYun Download Google Download
DAMO-YOLO-Ns 416 32.3 4.08 1.56 1.41 torch,onnx --
DAMO-YOLO-Nm 416 38.2 5.05 3.69 2.71 torch,onnx --
DAMO-YOLO-Nl 416 40.5 6.69 6.04 5.69 torch,onnx --
  • We report the mAP of models on COCO2017 validation set, with multi-class NMS.
  • The latency in this table is measured without post-processing, following picodet.
  • The latency is evaluated based on OpenVINO-2022.3.0, using commands below:
    # onnx export, enable --benchmark to ignore postprocess
    python tools/converter.py -f configs/damoyolo_tinynasL18_Ns.py -c ../damoyolo_tinynasL18_Ns.pth --batch_size 1  --img_size 416 --benchmark
    # model transform
    mo --input_model damoyolo_tinynasL18_Ns.onnx --data_type FP16
    # latency benchmark
    ./benchmark_app -m damoyolo_tinynasL18_Ns.xml -i ./assets/dog.jpg -api sync -d CPU -b 1 -hint latency 

701 categories DAMO-YOLO Model

We provide DAMO-YOLO-S model with 701 categories for general object detection, which has been trained on a large dataset including COCO, Objects365 and OpenImage. This model can also serve as a pre-trained model for fine-tuning in downstream tasks, enabling you to achieve better performance with ease.

Pretrained Model Downstream Task mAPval
0.5:0.95
AliYun Download Google Download
80-categories-DAMO-YOLO-S VisDrone 24.6 torch,onnx -
701-categories-DAMO-YOLO-S VisDrone 26.6 torch,onnx -
  • Note: The downloadable model is a pretrained model with 701 categories datasets. We demonstrate the VisDrone results to show that our pretrained model can enhance the performance of downstream tasks.

Quick Start

Installation

Step1. Install DAMO-YOLO.

git clone https://github.com/tinyvision/damo-yolo.git
cd DAMO-YOLO/
conda create -n DAMO-YOLO python=3.7 -y
conda activate DAMO-YOLO
conda install pytorch==1.7.0 torchvision==0.8.0 torchaudio==0.7.0 cudatoolkit=10.2 -c pytorch
pip install -r requirements.txt
export PYTHONPATH=$PWD:$PYTHONPATH

Step2. Install pycocotools.

pip install cython;
pip install git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI # for Linux
pip install git+https://github.com/philferriere/cocoapi.git#subdirectory=PythonAPI # for Windows
Demo

Step1. Download a pretrained torch, onnx or tensorRT engine from the benchmark table, e.g., damoyolo_tinynasL25_S.pth, damoyolo_tinynasL25_S.onnx, damoyolo_tinynasL25_S.trt.

Step2. Use -f(config filename) to specify your detector's config, --path to specify input data path, image/video/camera are supported. For example:

# torch engine with image
python tools/demo.py image -f ./configs/damoyolo_tinynasL25_S.py --engine ./damoyolo_tinynasL25_S.pth --conf 0.6 --infer_size 640 640 --device cuda --path ./assets/dog.jpg

# onnx engine with video
python tools/demo.py video -f ./configs/damoyolo_tinynasL25_S.py --engine ./damoyolo_tinynasL25_S.onnx --conf 0.6 --infer_size 640 640 --device cuda --path your_video.mp4

# tensorRT engine with camera
python tools/demo.py camera -f ./configs/damoyolo_tinynasL25_S.py --engine ./damoyolo_tinynasL25_S.trt --conf 0.6 --infer_size 640 640 --device cuda --camid 0
Reproduce our results on COCO

Step1. Prepare COCO dataset

cd <DAMO-YOLO Home>
ln -s /path/to/your/coco ./datasets/coco

Step 2. Reproduce our results on COCO by specifying -f(config filename)

python -m torch.distributed.launch --nproc_per_node=8 tools/train.py -f configs/damoyolo_tinynasL25_S.py
Finetune on your data

Please refer to custom dataset tutorial for details.

Evaluation
python -m torch.distributed.launch --nproc_per_node=8 tools/eval.py -f configs/damoyolo_tinynasL25_S.py --ckpt /path/to/your/damoyolo_tinynasL25_S.pth
Customize tinynas backbone Step1. If you want to customize your own backbone, please refer to [MAE-NAS Tutorial for DAMO-YOLO](https://github.com/alibaba/lightweight-neural-architecture-search/blob/main/scripts/damo-yolo/Tutorial_NAS_for_DAMO-YOLO_cn.md). This is a detailed tutorial about how to obtain an optimal backbone under the budget of latency/flops.

Step2. After the searching process completed, you can replace the structure text in configs with it. Finally, you can get your own custom ResNet-like or CSPNet-like backbone after setting the backbone name to TinyNAS_res or TinyNAS_csp. Please notice the difference of out_indices between TinyNAS_res and TinyNAS_csp.

structure = self.read_structure('tinynas_customize.txt')
TinyNAS = { 'name'='TinyNAS_res', # ResNet-like Tinynas backbone
            'out_indices': (2,4,5)}
TinyNAS = { 'name'='TinyNAS_csp', # CSPNet-like Tinynas backbone
            'out_indices': (2,3,4)}

Deploy

Installation

Step1. Install ONNX.

pip install onnx==1.8.1
pip install onnxruntime==1.8.0
pip install onnx-simplifier==0.3.5

Step2. Install CUDA、CuDNN、TensorRT and pyCUDA

2.1 CUDA

wget https://developer.download.nvidia.com/compute/cuda/10.2/Prod/local_installers/cuda_10.2.89_440.33.01_linux.run
sudo sh cuda_10.2.89_440.33.01_linux.run
export PATH=$PATH:/usr/local/cuda-10.2/bin
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda-10.2/lib64
source ~/.bashrc

2.2 CuDNN

sudo cp cuda/include/* /usr/local/cuda/include/
sudo cp cuda/lib64/libcudnn* /usr/local/cuda/lib64/
sudo chmod a+r /usr/local/cuda/include/cudnn.h
sudo chmod a+r /usr/local/cuda/lib64/libcudnn*

2.3 TensorRT

cd TensorRT-7.2.1.6/python
pip install tensorrt-7.2.1.6-cp37-none-linux_x86_64.whl
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:TensorRT-7.2.1.6/lib

2.4 pycuda

pip install pycuda==2022.1
Model Convert

Now we support trt_int8 quantization, you can specify trt_type as int8 to export the int8 tensorRT engine. You can also try partial quantization to achieve a good compromise between accuracy and latency. Refer to partial_quantization for more details.

Step.1 convert torch model to onnx or trt engine, and the output file would be generated in ./deploy. end2end means to export trt with nms. trt_eval means to evaluate the exported trt engine on coco_val dataset after the export compelete.

# onnx export 
python tools/converter.py -f configs/damoyolo_tinynasL25_S.py -c damoyolo_tinynasL25_S.pth --batch_size 1 --img_size 640

# trt export
python tools/converter.py -f configs/damoyolo_tinynasL25_S.py -c damoyolo_tinynasL25_S.pth --batch_size 1 --img_size 640 --trt --end2end --trt_eval

Step.2 trt engine evaluation on coco_val dataset. end2end means to using trt_with_nms to evaluation.

python tools/trt_eval.py -f configs/damoyolo_tinynasL25_S.py -trt deploy/damoyolo_tinynasL25_S_end2end_fp16_bs1.trt --batch_size 1 --img_size 640 --end2end

Step.3 onnx or trt engine inference demo and appoint test image/video by --path. end2end means to using trt_with_nms to inference.

# onnx inference
python tools/demo.py image -f ./configs/damoyolo_tinynasL25_S.py --engine ./damoyolo_tinynasL25_S.onnx --conf 0.6 --infer_size 640 640 --device cuda --path ./assets/dog.jpg

# trt inference
python tools/demo.py image -f ./configs/damoyolo_tinynasL25_S.py --engine ./deploy/damoyolo_tinynasL25_S_end2end_fp16_bs1.trt --conf 0.6 --infer_size 640 640 --device cuda --path ./assets/dog.jpg --end2end

Industry Application Models:

We provide DAMO-YOLO models for applications in real scenarios, which are listed as follows. More powerful models are coming, please stay tuned.

Human Detection Helmet Detection Head Detection Smartphone Detectioin
Facemask Detection Cigarette Detection Traffic Sign Detection NFL-helmet detection

Third Party Resources

In order to promote communication among DAMO-YOLO users, we collect third-party resources in this section. If you have original content about DAMO-YOLO, please feel free to contact us at xianzhe.xxz@alibaba-inc.com.

Cite DAMO-YOLO

If you use DAMO-YOLO in your research, please cite our work by using the following BibTeX entry:

 @article{damoyolo,
   title={DAMO-YOLO: A Report on Real-Time Object Detection Design},
   author={Xianzhe Xu, Yiqi Jiang, Weihua Chen, Yilun Huang, Yuan Zhang and Xiuyu Sun},
   journal={arXiv preprint arXiv:2211.15444v2},
   year={2022},
 }

 @inproceedings{sun2022mae,
   title={Mae-det: Revisiting maximum entropy principle in zero-shot nas for efficient object detection},
   author={Sun, Zhenhong and Lin, Ming and Sun, Xiuyu and Tan, Zhiyu and Li, Hao and Jin, Rong},
   booktitle={International Conference on Machine Learning},
   pages={20810--20826},
   year={2022},
   organization={PMLR}
 }

@inproceedings{jiang2022giraffedet,
  title={GiraffeDet: A Heavy-Neck Paradigm for Object Detection},
  author={yiqi jiang and Zhiyu Tan and Junyan Wang and Xiuyu Sun and Ming Lin and Hao Li},
  booktitle={International Conference on Learning Representations},
  year={2022},
}