EdgeYOLO: anchor-free, edge-friendly

简体中文

1 Intro
2 Updates
3 Coming Soon
4 Models
5 Quick Start
$\quad$5.1 setup
$\quad$5.2 inference
$\quad$5.3 train
$\quad$5.4 evaluate
$\quad$5.5 export onnx & tensorrt
6 Cite EdgeYOLO
7 Bugs found currently

Intro

In embeded device such as Nvidia Jetson AGX Xavier, EdgeYOLO reaches 34FPS with 50.6% AP in COCO2017 dataset and 25.9% AP in VisDrone2019 (image input size is 640x640, batch=16, post-process included). And for smaller model EdgeYOLO-S, it reaches 53FPS with 44.1% AP and 63.3% AP^0.5(SOTA in P5 small models) in COCO2017.
we provide a more effective data augmentation during training.
small object and medium object detect performace is imporved by using RH loss during the last few training epochs.
Our pre-print paper is released on arxiv.

Updates

[2023/2/20]

TensorRT cpp inference console demo (lib opencv and qt5 required)
Fix bugs when exporting models using Version 7.X TensorRT

[2023/2/19]

Publish TensorRT int8 export code with Calibration (torch2trt is required)

Coming Soon

Now evaluate.py doesn't support tensorrt model, we will update it in the near future
MNN deployment code
More different models
C++ code for TensorRT inference with UI
EdgeYOLO-mask for segmentation task
Simple but effective pretrain method

Models

models trained on COCO2017-train

Model	Size	mAP^val 0.5:0.95	mAP^val 0.5	FPS^{AGX Xavier trt fp16 batch=16 include NMS}	Params train / infer ^(M)	Download
EdgeYOLO-Tiny-LRELU	416 640	33.1 37.8	50.5 56.7	206 109	7.6 / 7.0	github
EdgeYOLO-Tiny	416 640	37.2 41.4	55.4 60.4	136 67	5.8 / 5.5	github
EdgeYOLO-S	640	44.1	63.3	53	9.9 / 9.3	github
EdgeYOLO-M	640	47.5	66.6	46	19.0 / 17.8	github
EdgeYOLO	640	50.6	69.8	34	41.2 / 40.5	github

models trained on VisDrone2019 (pretrained backbone on COCO2017-train)

we use VisDrone2019-DET dataset with COCO format in our training.

Model	Size	mAP^val 0.5:0.95	mAP^val 0.5	Download
EdgeYOLO-Tiny-LRELU	416 640	12.1 18.5	22.8 33.6	github
EdgeYOLO-Tiny	416 640	14.9 21.8	27.3 38.5	github
EdgeYOLO-S	640	23.6	40.8	github
EdgeYOLO-M	640	25.0	42.9	github
EdgeYOLO	640	25.9 26.4	43.9 44.8	github(legacy) github(new)

Some of our detect results in COCO2017

COCO2017

Quick Start

setup

git clone https://github.com/LSH9832/edgeyolo.git
cd edgeyolo
pip install -r requirements.txt

if you use tensorrt, please make sure torch2trt and TensorRT Development Toolkit(version>7.1.3.0) are installed.

git clone https://github.com/NVIDIA-AI-IOT/torch2trt.git
cd torch2trt
python setup.py install

or to make sure you use the same version of torch2trt as ours, download here

inference

First download weights here

python detect.py --weights edgeyolo_coco.pth --source XXX.mp4 --fp16

# all options
python detect.py --weights edgeyolo_coco.pth 
                 --source /XX/XXX.mp4     # or dir with images, such as /dataset/coco2017/val2017    (jpg/jpeg, png, bmp, webp is available)
                 --conf-thres 0.25 
                 --nms-thres 0.5 
                 --input-size 640 640 
                 --batch 1 
                 --save-dir ./output/detect/imgs    # if you press "s", the current frame will be saved in this dir
                 --fp16 
                 --no-fuse                # do not reparameterize model
                 --no-label               # do not draw label with class name and confidence
                 --mp                     # use multi-process to show images more smoothly when batch > 1
                 --fps 30                 # max fps limitation, valid only when option --mp is used

train

first preparing your dataset and create dataset config file(./params/dataset/XXX.yaml), make sure your dataset config file contains:

(COCO, VOC, VisDrone and DOTA formats are supported)

type: "coco"                        # dataset format(lowercase)，COCO, VOC, VisDrone and DOTA formats are supported currently
dataset_path: "/dataset/coco2017"   # root dir of your dataset

kwargs:
  suffix: "jpg"        # suffix of your dataset's images
  use_cache: true      # test on i5-12490f: Total loading time: 52s -> 10s(seg enabled) and 39s -> 4s(seg disabled)

train:
  image_dir: "images/train2017"                   # train set image dir
  label: "annotations/instances_train2017.json"   # train set label file(format with single label file) or directory(multi label files)

val:
  image_dir: "images/val2017"                     # evaluate set image dir
  label: "annotations/instances_val2017.json"     # evaluate set label file or directory

test:
  test_dir: "test2017"     # test set image dir (not used in code now, but will)

segmentaion_enabled: true  # whether this dataset has segmentation labels and you are going to use them instead of bbox labels

names: ['person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light',
        'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow',
        'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee',
        'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard',
        'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple',
        'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch',
        'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone',
        'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear',
        'hair drier', 'toothbrush']    # category names

then edit file ./params/train/train_XXX.yaml
finally

python train.py --cfg ./params/train/train_XXX.yaml

evaluate

python evaluate.py --weights edgeyolo_coco.pth --dataset params/dataset/XXX.yaml --batch 16 --device 0

# all options
python evaluate.py --weights edgeyolo_coco.pth 
                   --dataset params/dataset/XXX.yaml 
                   --batch 16   # batch size for each gpu
                   --device 0
                   --input-size 640 640   # height, width

export onnx & tensorrt

ONNX

python export.py --onnx --weights edgeyolo_coco.pth --batch 1

# all options
python export.py --onnx   # or --onnx-only if tensorrt and torch2trt are not installed
                 --weights edgeyolo_coco.pth 
                 --input-size 640 640   # height, width
                 --batch 1
                 --opset 11
                 --no-simplify    # do not simplify this model

it generates

output/export/edgeyolo_coco/640x640_batch1.onnx

TensorRT

# fp16
python export.py --trt --weights edgeyolo_coco.pth --batch 1 --workspace 8

# int8
python export.py --trt --weights edgeyolo_coco.pth --batch 1 --workspace 8 --int8 --dataset params/dataset/coco.yaml --num-imgs 1024

# all options
python export.py --trt                       # you can add --onnx and relative options to export both models
                 --weights edgeyolo_coco.pth
                 --input-size 640 640        # height, width
                 --batch 1
                 --workspace 10              # (GB)
                 --no-fp16        # fp16 mode in default, use this option to disable it(fp32)
                 --int8           # int8 mode, the following options are needed for calibration
                 --datset params/dataset/coco.yaml   # generates calibration images from its val images(upper limit：5120)
                 --train          # use train images instead of val images(upper limit：5120)
                 --all            # use all images(upper limit：5120)
                 --num-imgs 512   # (upper limit：5120)

it generates

(optional) output/export/edgeyolo_coco/640x640_batch1.onnx
output/export/edgeyolo_coco/640x640_batch1_fp16/int8.pt       # for python inference
output/export/edgeyolo_coco/640x640_batch1_fp16/int8.engine   # for c++ inference
output/export/edgeyolo_coco/640x640_batch1_fp16/int8.json     # for c++ inference

Benchmark of TensorRT Int8 Model

enviroment: TensorRT Version 8.2.5.1, Windows, i5-12490F, RTX 3060 12GB
increase workspace and the number of images for calibration may improve the performance

COCO2017-TensorRT-int8

Int8 Model	Size	Calibration Image number	Workspace ^(GB)	mAP^val 0.5:0.95	mAP^val 0.5	FPS^{RTX 3060 trt int8 batch=16 include NMS}
Tiny-LRELU	416 640	512	8	31.5 36.4	48.7 55.5	730 360
Tiny	416 640	512	8	34.9 39.8	53.1 59.5	549 288
S	640	512	8	42.4	61.8	233
M	640	512	8	45.2	64.2	211
L	640	512	8	49.1	68.0	176

for python inference

python detect.py --trt --weights output/export/edgeyolo_coco/640x640_batch1_int8.pt --source XXX.mp4

# all options
python detect.py --trt 
                 --weights output/export/edgeyolo_coco/640x640_batch1_int8.pt 
                 --source XXX.mp4
                 --legacy         # if "img = img / 255" when you train your train model
                 --use-decoder    # if use original yolox tensorrt model before version 0.3.0
                 --mp             # use multi-process to show images more smoothly when batch > 1
                 --fps 30         # max fps limitation, valid only when option --mp is used

for c++ inference

# build
cd cpp/console/linux
mkdir build && cd build
cmake ..
make -j4

# help
./yolo -?
./yolo --help

# run
# ./yolo [json file] [source] [--conf] [--nms] [--loop] [--no-label]
./yolo ../../../../output/export/edgeyolo_coco/640x640_batch1_int8.json ~/Videos/test.avi --conf 0.25 --nms 0.5 --loop --no-label

Cite EdgeYOLO

 @article{edgeyolo2023,
  title={EdgeYOLO: An Edge-Real-Time Object Detector},
  author={Shihan Liu, Junlin Zha, Jian Sun, Zhuo Li, and Gang Wang},
  journal={arXiv preprint arXiv:2302.07483},
  year={2023}
}

Bugs found currently

Sometimes it raises error as follows during training. Reduce pytorch version to 1.8.0 might solve this problem.

File "XXX/edgeyolo/edgeyolo/train/loss.py", line 667, in dynamic_k_matching
_, pos_idx = torch.topk(cost[gt_idx], k=dynamic_ks[gt_idx].item(), largest=False)
RuntimeError: CUDA error: device-side assert triggered
CUDA kernel errors might be asynchronously reported at some other API call,so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1.

For DOTA dataset, we only support single GPU training mode now, please do not train DOTA dataset with distributed mode or model can not be trained correctly.
Sometimes converting to TensorRT fp16 model with 8.4.X.X or higher version might lose a lot of precision, please use TensorRT Verson 7.X.X.X or 8.2.X.X

lbwzv / edgeyolo Goto Github PK

edgeyolo's Introduction