Quantization is all you need.

YOLOv6

Implementation of paper:

What's New

[2023.09.15] Release YOLOv6-Segmentation. 🚀 Performance
[2023.04.28] Release YOLOv6Lite models on mobile or CPU. ⭐️ Mobile Benchmark
[2023.03.10] Release YOLOv6-Face. 🔥 Performance
[2023.03.02] Update base models to version 3.0.
[2023.01.06] Release P6 models and enhance the performance of P5 models. ⭐️ Benchmark
[2022.11.04] Release base models to simplify the training and deployment process.
[2022.09.06] Customized quantization methods. 🚀 Quantization Tutorial
[2022.09.05] Release M/L models and update N/T/S models with enhanced performance.
[2022.06.23] Release N/T/S models with excellent performance.

Benchmark

Model	Size	mAP^val 0.5:0.95	Speed^{T4 trt fp16 b1 (fps)}	Speed^{T4 trt fp16 b32 (fps)}	Params ^(M)	FLOPs ^(G)
YOLOv6-N	640	37.5	779	1187	4.7	11.4
YOLOv6-S	640	45.0	339	484	18.5	45.3
YOLOv6-M	640	50.0	175	226	34.9	85.8
YOLOv6-L	640	52.8	98	116	59.6	150.7

YOLOv6-N6	1280	44.9	228	281	10.4	49.8
YOLOv6-S6	1280	50.3	98	108	41.4	198.0
YOLOv6-M6	1280	55.2	47	55	79.6	379.5
YOLOv6-L6	1280	57.2	26	29	140.4	673.4

Table Notes

All checkpoints are trained with self-distillation except for YOLOv6-N6/S6 models trained to 300 epochs without distillation.
Results of the mAP and speed are evaluated on COCO val2017 dataset with the input resolution of 640×640 for P5 models and 1280x1280 for P6 models.
Speed is tested with TensorRT 7.2 on T4.
Refer to Test speed tutorial to reproduce the speed results of YOLOv6.
Params and FLOPs of YOLOv6 are estimated on deployed models.

Legacy models

Model	Size	mAP^val 0.5:0.95	Speed^{T4 trt fp16 b1 (fps)}	Speed^{T4 trt fp16 b32 (fps)}	Params ^(M)	FLOPs ^(G)
YOLOv6-N	640	35.9^300e 36.3^400e	802	1234	4.3	11.1
YOLOv6-T	640	40.3^300e 41.1^400e	449	659	15.0	36.7
YOLOv6-S	640	43.5^300e 43.8^400e	358	495	17.2	44.2
YOLOv6-M	640	49.5	179	233	34.3	82.2
YOLOv6-L-ReLU	640	51.7	113	149	58.5	144.0
YOLOv6-L	640	52.5	98	121	58.5	144.0

Speed is tested with TensorRT 7.2 on T4.

Quantized model 🚀

Model	Size	Precision	mAP^val 0.5:0.95	Speed^{T4 trt b1 (fps)}	Speed^{T4 trt b32 (fps)}
YOLOv6-N RepOpt	640	INT8	34.8	1114	1828
YOLOv6-N	640	FP16	35.9	802	1234
YOLOv6-T RepOpt	640	INT8	39.8	741	1167
YOLOv6-T	640	FP16	40.3	449	659
YOLOv6-S RepOpt	640	INT8	43.3	619	924
YOLOv6-S	640	FP16	43.5	377	541

Speed is tested with TensorRT 8.4 on T4.
Precision is figured on models for 300 epochs.

Mobile Benchmark

Model	Size	mAP^val 0.5:0.95	sm8350 ^(ms)	mt6853 ^(ms)	sdm660 ^(ms)	Params ^(M)	FLOPs ^(G)
YOLOv6Lite-S	320*320	22.4	7.99	11.99	41.86	0.55	0.56
YOLOv6Lite-M	320*320	25.1	9.08	13.27	47.95	0.79	0.67
YOLOv6Lite-L	320*320	28.0	11.37	16.20	61.40	1.09	0.87
YOLOv6Lite-L	320*192	25.0	7.02	9.66	36.13	1.09	0.52
YOLOv6Lite-L	224*128	18.9	3.63	4.99	17.76	1.09	0.24

Table Notes

From the perspective of model size and input image ratio, we have built a series of models on the mobile terminal to facilitate flexible applications in different scenarios.
All checkpoints are trained with 400 epochs without distillation.
Results of the mAP and speed are evaluated on COCO val2017 dataset, and the input resolution is the Size in the table.
Speed is tested on MNN 2.3.0 AArch64 with 2 threads by arm82 acceleration. The inference warm-up is performed 10 times, and the cycle is performed 100 times.
Qualcomm 888(sm8350), Dimensity 720(mt6853) and Qualcomm 660(sdm660) correspond to chips with different performances at the high, middle and low end respectively, which can be used as a reference for model capabilities under different chips.
Refer to Test NCNN Speed tutorial to reproduce the NCNN speed results of YOLOv6Lite.

Quick Start

Install

git clone https://github.com/meituan/YOLOv6
cd YOLOv6
pip install -r requirements.txt

Reproduce our results on COCO

Please refer to Train COCO Dataset.

Finetune on custom data

Single GPU

# P5 models
python tools/train.py --batch 32 --conf configs/yolov6s_finetune.py --data data/dataset.yaml --fuse_ab --device 0
# P6 models
python tools/train.py --batch 32 --conf configs/yolov6s6_finetune.py --data data/dataset.yaml --img 1280 --device 0

Multi GPUs (DDP mode recommended)

# P5 models
python -m torch.distributed.launch --nproc_per_node 8 tools/train.py --batch 256 --conf configs/yolov6s_finetune.py --data data/dataset.yaml --fuse_ab --device 0,1,2,3,4,5,6,7
# P6 models
python -m torch.distributed.launch --nproc_per_node 8 tools/train.py --batch 128 --conf configs/yolov6s6_finetune.py --data data/dataset.yaml --img 1280 --device 0,1,2,3,4,5,6,7

fuse_ab: add anchor-based auxiliary branch and use Anchor Aided Training Mode (Not supported on P6 models currently)
conf: select config file to specify network/optimizer/hyperparameters. We recommend to apply yolov6n/s/m/l_finetune.py when training on your custom dataset.
data: prepare dataset and specify dataset paths in data.yaml ( COCO, YOLO format coco labels )
make sure your dataset structure as follows:

├── coco
│   ├── annotations
│   │   ├── instances_train2017.json
│   │   └── instances_val2017.json
│   ├── images
│   │   ├── train2017
│   │   └── val2017
│   ├── labels
│   │   ├── train2017
│   │   ├── val2017
│   ├── LICENSE
│   ├── README.txt

YOLOv6 supports different input resolution modes. For details, see How to Set the Input Size.

Resume training

If your training process is corrupted, you can resume training by

# single GPU training.
python tools/train.py --resume

# multi GPU training.
python -m torch.distributed.launch --nproc_per_node 8 tools/train.py --resume

Above command will automatically find the latest checkpoint in YOLOv6 directory, then resume the training process.

Your can also specify a checkpoint path to --resume parameter by

# remember to replace /path/to/your/checkpoint/path to the checkpoint path which you want to resume training.
--resume /path/to/your/checkpoint/path

This will resume from the specific checkpoint you provide.

Evaluation

Reproduce mAP on COCO val2017 dataset with 640×640 or 1280x1280 resolution

# P5 models
python tools/eval.py --data data/coco.yaml --batch 32 --weights yolov6s.pt --task val --reproduce_640_eval
# P6 models
python tools/eval.py --data data/coco.yaml --batch 32 --weights yolov6s6.pt --task val --reproduce_640_eval --img 1280

verbose: set True to print mAP of each classes.
do_coco_metric: set True / False to enable / disable pycocotools evaluation method.
do_pr_metric: set True / False to print or not to print the precision and recall metrics.
config-file: specify a config file to define all the eval params, for example: yolov6n_with_eval_params.py

Inference

First, download a pretrained model from the YOLOv6 release or use your trained model to do inference.

Second, run inference with tools/infer.py

# P5 models
python tools/infer.py --weights yolov6s.pt --source img.jpg / imgdir / video.mp4
# P6 models
python tools/infer.py --weights yolov6s6.pt --img 1280 1280 --source img.jpg / imgdir / video.mp4

If you want to inference on local camera or web camera, you can run:

# P5 models
python tools/infer.py --weights yolov6s.pt --webcam --webcam-addr 0
# P6 models
python tools/infer.py --weights yolov6s6.pt --img 1280 1280 --webcam --webcam-addr 0

webcam-addr can be local camera number id or rtsp address.

Deployment

Tutorials

Third-party resources