YOLOv5-Lite：Lighter, faster and easier to deploy

Perform a series of ablation experiments on yolov5 to make it lighter (smaller Flops, lower memory, and fewer parameters) and faster (add shuffle channel, yolov5 head for channel reduce. It can infer at least 10+ FPS On the Raspberry Pi 4B when input the frame with 320×320) and is easier to deploy (removing the Focus layer and four slice operations, reducing the model quantization accuracy to an acceptable range).

Comparison of ablation experiment results

ID	Model	Input_size	Flops	Params	Size（M）	[email protected]	[email protected]:0.95
001	yolo-fastest	320×320	0.25G	0.35M	1.4	24.4	-
002	YOLOv5-Liteeours	320×320	0.73G	0.78M	1.7	35.1	-
003	NanoDet-m	320×320	0.72G	0.95M	1.8	-	20.6
004	yolo-fastest-xl	320×320	0.72G	0.92M	3.5	34.3	-
005	YOLOXNano	416×416	1.08G	0.91M	7.3(fp32)	-	25.8
006	yolov3-tiny	416×416	6.96G	6.06M	23.0	33.1	16.6
007	yolov4-tiny	416×416	5.62G	8.86M	33.7	40.2	21.7
008	YOLOv5-Litesours	416×416	1.66G	1.64M	3.4	42.0	25.2
009	YOLOv5-Litecours	512×512	5.92G	4.57M	9.2	50.9	32.5
010	NanoDet-EfficientLite2	512×512	7.12G	4.71M	18.3	-	32.6
011	YOLOv5s(6.0)	640×640	16.5G	7.23M	14.0	56.0	37.2
012	YOLOv5-Litegours	640×640	15.6G	5.39M	10.9	57.6	39.1

See the wiki: https://github.com/ppogg/YOLOv5-Lite/wiki/Test-the-map-of-models-about-coco

Comparison on different platforms

Equipment	Computing backend	System	Input	Framework	v5lite-e	v5lite-s	v5lite-c	v5lite-g	YOLOv5s
Inter	@i5-10210U	window(x86)	640×640	openvino	-	-	46ms	-	131ms
Nvidia	@RTX 2080Ti	Linux(x86)	640×640	torch	-	-	-	15ms	14ms
Redmi K30	@Snapdragon 730G	Android(armv8)	320×320	ncnn	27ms	38ms	-	-	163ms
Xiaomi 10	@Snapdragon 865	Android(armv8)	320×320	ncnn	10ms	14ms	-	-	163ms
Raspberrypi 4B	@ARM Cortex-A72	Linux(arm64)	320×320	ncnn	-	84ms	-	-	371ms
Raspberrypi 4B	@ARM Cortex-A72	Linux(arm64)	320×320	mnn	-	71ms	-	-	356ms
AXera-Pi	Cortex A7@CPU 3.6TOPs @NPU	Linux(arm64)	640×640	axpi	-	-	-	22ms	22ms

The tutorial of 15FPS on Raspberry Pi 4B:

https://zhuanlan.zhihu.com/p/672633849

• The above is a 4-thread test benchmark
• Raspberrypi 4B enable bf16s optimization，Raspberrypi 64 Bit OS

qq交流群：993965802

入群答案:剪枝 or 蒸馏 or 量化 or 低秩分解（任意其一均可）

·Model Zoo·

@v5lite-e:

Model	Size	Backbone	Head	Framework	Design for
v5Lite-e.pt	1.7m	shufflenetv2（Megvii）	v5Litee-head	Pytorch	Arm-cpu
v5Lite-e.bin v5Lite-e.param	1.7m	shufflenetv2	v5Litee-head	ncnn	Arm-cpu
v5Lite-e-int8.bin v5Lite-e-int8.param	0.9m	shufflenetv2	v5Litee-head	ncnn	Arm-cpu
v5Lite-e-fp32.mnn	3.0m	shufflenetv2	v5Litee-head	mnn	Arm-cpu
v5Lite-e-fp32.tnnmodel v5Lite-e-fp32.tnnproto	2.9m	shufflenetv2	v5Litee-head	tnn	arm-cpu
v5Lite-e-320.onnx	3.1m	shufflenetv2	v5Litee-head	onnxruntime	x86-cpu

@v5lite-s:

Model	Size	Backbone	Head	Framework	Design for
v5Lite-s.pt	3.4m	shufflenetv2（Megvii）	v5Lites-head	Pytorch	Arm-cpu
v5Lite-s.bin v5Lite-s.param	3.3m	shufflenetv2	v5Lites-head	ncnn	Arm-cpu
v5Lite-s-int8.bin v5Lite-s-int8.param	1.7m	shufflenetv2	v5Lites-head	ncnn	Arm-cpu
v5Lite-s.mnn	3.3m	shufflenetv2	v5Lites-head	mnn	Arm-cpu
v5Lite-s-int4.mnn	987k	shufflenetv2	v5Lites-head	mnn	Arm-cpu
v5Lite-s-fp16.bin v5Lite-s-fp16.xml	3.4m	shufflenetv2	v5Lites-head	openvivo	x86-cpu
v5Lite-s-fp32.bin v5Lite-s-fp32.xml	6.8m	shufflenetv2	v5Lites-head	openvivo	x86-cpu
v5Lite-s-fp16.tflite	3.3m	shufflenetv2	v5Lites-head	tflite	arm-cpu
v5Lite-s-fp32.tflite	6.7m	shufflenetv2	v5Lites-head	tflite	arm-cpu
v5Lite-s-int8.tflite	1.8m	shufflenetv2	v5Lites-head	tflite	arm-cpu
v5Lite-s-416.onnx	6.4m	shufflenetv2	v5Lites-head	onnxruntime	x86-cpu

@v5lite-c:

Model	Size	Backbone	Head	Framework	Design for
v5Lite-c.pt	9m	PPLcnet（Baidu）	v5s-head	Pytorch	x86-cpu / x86-vpu
v5Lite-c.bin v5Lite-c.xml	8.7m	PPLcnet	v5s-head	openvivo	x86-cpu / x86-vpu
v5Lite-c-512.onnx	18m	PPLcnet	v5s-head	onnxruntime	x86-cpu

@v5lite-g:

Model	Size	Backbone	Head	Framework	Design for
v5Lite-g.pt	10.9m	Repvgg（Tsinghua）	v5Liteg-head	Pytorch	x86-gpu / arm-gpu / arm-npu
v5Lite-g-int8.engine	8.5m	Repvgg-yolov5	v5Liteg-head	Tensorrt	x86-gpu / arm-gpu / arm-npu
v5lite-g-int8.tmfile	8.7m	Repvgg-yolov5	v5Liteg-head	Tengine	arm-npu
v5Lite-g-640.onnx	21m	Repvgg-yolov5	yolov5-head	onnxruntime	x86-cpu
v5Lite-g-640.joint	7.1m	Repvgg-yolov5	yolov5-head	axpi	arm-npu

Download Link：

• v5lite-e.pt: | Baidu Drive | Google Drive |
  

|──────ncnn-fp16: | Baidu Drive | Google Drive |
|──────ncnn-int8: | Baidu Drive | Google Drive |
|──────mnn-e_bf16: | Google Drive |
|──────mnn-d_bf16: | Google Drive|
└──────onnx-fp32: | Baidu Drive | Google Drive |

• v5lite-s.pt: | Baidu Drive | Google Drive |
  

|──────ncnn-fp16: | Baidu Drive | Google Drive |
|──────ncnn-int8: | Baidu Drive | Google Drive |
└──────tengine-fp32: | Baidu Drive | Google Drive |

• v5lite-c.pt: Baidu Drive | Google Drive |
  

└──────openvino-fp16: | Baidu Drive | Google Drive |

• v5lite-g.pt: | Baidu Drive | Google Drive |
  

└──────axpi-int8: Google Drive |

Baidu Drive Password: pogg

v5lite-s model: TFLite Float32, Float16, INT8, Dynamic range quantization, ONNX, TFJS, TensorRT, OpenVINO IR FP32/FP16, Myriad Inference Engin Blob, CoreML

https://github.com/PINTO0309/PINTO_model_zoo/tree/main/180_YOLOv5-Lite

Thanks for PINTO0309:https://github.com/PINTO0309/PINTO_model_zoo/tree/main/180_YOLOv5-Lite

How to use

Install

Python>=3.6.0 is required with all requirements.txt installed including PyTorch>=1.7:

$ git clone https://github.com/ppogg/YOLOv5-Lite
$ cd YOLOv5-Lite
$ pip install -r requirements.txt

Inference with detect.py

detect.py runs inference on a variety of sources, downloading models automatically from the latest YOLOv5-Lite release and saving results to runs/detect.

$ python detect.py --source 0  # webcam
                            file.jpg  # image 
                            file.mp4  # video
                            path/  # directory
                            path/*.jpg  # glob
                            'https://youtu.be/NUsoVlDFqZg'  # YouTube
                            'rtsp://example.com/media.mp4'  # RTSP, RTMP, HTTP stream

Training

$ python train.py --data coco.yaml --cfg v5lite-e.yaml --weights v5lite-e.pt --batch-size 128
                                         v5lite-s.yaml           v5lite-s.pt              128
                                         v5lite-c.yaml           v5lite-c.pt               96
                                         v5lite-g.yaml           v5lite-g.pt               64

If you use multi-gpu. It's faster several times:

$ python -m torch.distributed.launch --nproc_per_node 2 train.py

DataSet

Training set and test set distribution （the path with xx.jpg）

train: ../coco/images/train2017/
val: ../coco/images/val2017/

├── images            # xx.jpg example
│   ├── train2017        
│   │   ├── 000001.jpg
│   │   ├── 000002.jpg
│   │   └── 000003.jpg
│   └── val2017         
│       ├── 100001.jpg
│       ├── 100002.jpg
│       └── 100003.jpg
└── labels             # xx.txt example      
    ├── train2017       
    │   ├── 000001.txt
    │   ├── 000002.txt
    │   └── 000003.txt
    └── val2017         
        ├── 100001.txt
        ├── 100002.txt
        └── 100003.txt

Auto LabelImg

Link ：https://github.com/ppogg/AutoLabelImg

You can use LabelImg based YOLOv5-5.0 and YOLOv5-Lite to AutoAnnotate, biubiubiu 🚀 🚀 🚀

Model Hub

Here, the original components of YOLOv5 and the reproduced components of YOLOv5-Lite are organized and stored in the model hub：

Heatmap Analysis

$ python main.py --type all

Updating ...

How to deploy

ncnn for arm-cpu

mnn for arm-cpu

openvino x86-cpu or x86-vpu

tensorrt(C++) for arm-gpu or arm-npu or x86-gpu

tensorrt(Python) for arm-gpu or arm-npu or x86-gpu

Android for arm-cpu

Android_demo

This is a Redmi phone, the processor is Snapdragon 730G, and yolov5-lite is used for detection. The performance is as follows:

link: https://github.com/ppogg/YOLOv5-Lite/tree/master/android_demo/ncnn-android-v5lite

Android_v5Lite-s: https://drive.google.com/file/d/1CtohY68N2B9XYuqFLiTp-Nd2kuFWgAUR/view?usp=sharing

Android_v5Lite-g: https://drive.google.com/file/d/1FnvkWxxP_aZwhi000xjIuhJ_OhqOUJcj/view?usp=sharing

new android app:[link] https://pan.baidu.com/s/1PRhW4fI1jq8VboPyishcIQ [keyword] pogg

More detailed explanation

Detailed model link:

What is YOLOv5-Lite S/E model: zhihu link (Chinese): https://zhuanlan.zhihu.com/p/400545131

What is YOLOv5-Lite C model: zhihu link (Chinese): https://zhuanlan.zhihu.com/p/420737659

What is YOLOv5-Lite G model: zhihu link (Chinese): https://zhuanlan.zhihu.com/p/410874403

How to deploy on ncnn with fp16 or int8: csdn link (Chinese): https://blog.csdn.net/weixin_45829462/article/details/119787840

How to deploy on mnn with fp16 or int8: zhihu link (Chinese): https://zhuanlan.zhihu.com/p/672633849

How to deploy on onnxruntime: zhihu link (Chinese): https://zhuanlan.zhihu.com/p/476533259(old version)

How to deploy on tensorrt: zhihu link (Chinese): https://zhuanlan.zhihu.com/p/478630138

How to optimize on tensorrt: zhihu link (Chinese): https://zhuanlan.zhihu.com/p/463074494

Reference

https://github.com/ultralytics/yolov5

https://github.com/megvii-model/ShuffleNet-Series

https://github.com/Tencent/ncnn

Citing YOLOv5-Lite

If you use YOLOv5-Lite in your research, please cite our work and give a star ⭐:

 @misc{yolov5lite2021,
  title = {YOLOv5-Lite: Lighter, faster and easier to deploy},
  author = {Xiangrong Chen and Ziman Gong},
  doi = {10.5281/zenodo.5241425}
  year={2021}
}