English | 简体中文

YOLOv6

Implementation of paper - YOLOv6: A Single-Stage Object Detection Framework for Industrial Applications

YOLOv6 has a series of models for various industrial scenarios, including N/T/S/M/L, which the architectures vary considering the model size for better accuracy-speed trade-off. And some Bag-of-freebies methods are introduced to further improve the performance, such as self-distillation and more training epochs. For industrial deployment, we adopt QAT with channel-wise distillation and graph optimization to pursue extreme performance.

YOLOv6-N hits 35.9% AP on COCO dataset with 1234 FPS on T4. YOLOv6-S strikes 43.5% AP with 495 FPS, and the quantized YOLOv6-S model achieves 43.3% AP at a accelerated speed of 869 FPS on T4. YOLOv6-T/M/L also have excellent performance, which show higher accuracy than other detectors with the similar inference speed.

What's New

• [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.⭐️ Benchmark
• [2022.06.23] Release N/T/S models with excellent performance.

Benchmark

Model	Size	mAPval 0.5:0.95	SpeedT4 trt fp16 b1 (fps)	SpeedT4 trt fp16 b32 (fps)	Params (M)	FLOPs (G)
YOLOv6-N	640	35.9300e 36.3400e	802	1234	4.3	11.1
YOLOv6-T	640	40.3300e 41.1400e	449	659	9.7	24.9
YOLOv6-S	640	43.5300e 43.8400e	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	mAPval 0.5:0.95	SpeedT4 trt b1 (fps)	SpeedT4 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.
• Results of the mAP and speed are evaluated on COCO val2017 dataset with the input resolution of 640×640.
• 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	mAPval 0.5:0.95	SpeedV100 fp16 b32 (ms)	SpeedV100 fp32 b32 (ms)	SpeedT4 trt fp16 b1 (fps)	SpeedT4 trt fp16 b32 (fps)	Params (M)	FLOPs (G)
YOLOv6-N	416 640	30.8 35.0	0.3 0.5	0.4 0.7	1100 788	2716 1242	4.3 4.3	4.7 11.1
YOLOv6-T	640	41.3	0.9	1.5	425	602	15.0	36.7
YOLOv6-S	640	43.1	1.0	1.7	373	520	17.2	44.2

Quick Start

Install

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

Training

Single GPU

python tools/train.py --batch 32 --conf configs/yolov6s_finetune.py --data data/dataset.yaml --device 0

Multi GPUs (DDP mode recommended)

python -m torch.distributed.launch --nproc_per_node 8 tools/train.py --batch 256 --conf configs/yolov6s_finetune.py --data data/dataset.yaml --device 0,1,2,3,4,5,6,7

• 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 COCO dataset, YOLO format coco labels and specify dataset paths in data.yaml
• make sure your dataset structure as follows:

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

Reproduce our results on COCO ⭐️

For nano model

python -m torch.distributed.launch --nproc_per_node 4 tools/train.py \
									--batch 128 \
									--conf configs/yolov6n.py \
									--data data/coco.yaml \
									--epoch 400 \
									--device 0,1,2,3 \
									--name yolov6n_coco

For s/tiny model

python -m torch.distributed.launch --nproc_per_node 8 tools/train.py \
									--batch 256 \
									--conf configs/yolov6s.py \ # configs/yolov6t.py
									--data data/coco.yaml \
									--epoch 400 \
									--device 0,1,2,3,4,5,6,7 \
									--name yolov6s_coco # yolov6t_coco

For m/l model

# Step 1: Training a base model
python -m torch.distributed.launch --nproc_per_node 8 tools/train.py \
									--batch 256 \
									--conf configs/yolov6m.py \ # configs/yolov6l.py
									--data data/coco.yaml \
									--epoch 300 \
									--device 0,1,2,3,4,5,6,7 \
									--name yolov6m_coco # yolov6l_coco


# Step 2: Self-distillation training
python -m torch.distributed.launch --nproc_per_node 8 tools/train.py \
									--batch 256 \ # 128 for distillation of yolov6l
									--conf configs/yolov6m.py \ # configs/yolov6l.py
									--data data/coco.yaml \
									--epoch 300 \
									--device 0,1,2,3,4,5,6,7 \
									--distill \
									--teacher_model_path runs/train/yolov6m_coco/weights/best_ckpt.pt \ # # yolov6l_coco
									--name yolov6m_coco # yolov6l_coco

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 resolution ⭐️

python tools/eval.py --data data/coco.yaml --batch 32 --weights yolov6s.pt --task val --reproduce_640_eval

• 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

python tools/infer.py --weights yolov6s.pt --source img.jpg / imgdir / video.mp4

Deployment

• ONNX
• OpenCV Python/C++
• OpenVINO
• TensorRT

Tutorials

• Train custom data
• Test speed
• Tutorial of Quantization for YOLOv6

Third-party resources

• YOLOv6 NCNN Android app demo: ncnn-android-yolov6 from FeiGeChuanShu

• YOLOv6 ONNXRuntime/MNN/TNN C++: YOLOv6-ORT, YOLOv6-MNN and YOLOv6-TNN from DefTruth

• YOLOv6 TensorRT Python: yolov6-tensorrt-python from Linaom1214

• YOLOv6 TensorRT Windows C++: yolort from Wei Zeng

• YOLOv6 web demo on Huggingface Spaces with Gradio.

• Tutorial: How to train YOLOv6 on a custom dataset

• YouTube Tutorial: How to train YOLOv6 on a custom dataset

• Demo of YOLOv6 inference on Google Colab

• Blog post: YOLOv6 Object Detection – Paper Explanation and Inference

FAQ（Continuously updated）