English | 简体中文

PP-YOLO

Table of Contents

• Introduction
• Model Zoo
• Getting Start
• Future Work
• Appendix

Introduction

PP-YOLO is a optimized model based on YOLOv3 in PaddleDetection，whose performance(mAP on COCO) and inference speed are better than YOLOv4，PaddlePaddle 2.0.0rc1(available on pip now) or Daily Version is required to run this PP-YOLO。

PP-YOLO reached mAP(IoU=0.5:0.95) as 45.9% on COCO test-dev2017 dataset, and inference speed of FP32 on single V100 is 72.9 FPS, inference speed of FP16 with TensorRT on single V100 is 155.6 FPS.

PP-YOLO improved performance and speed of YOLOv3 with following methods:

• Better backbone: ResNet50vd-DCN
• Larger training batch size: 8 GPUs and mini-batch size as 24 on each GPU
• Drop Block
• Exponential Moving Average
• IoU Loss
• Grid Sensitive
• Matrix NMS
• CoordConv
• Spatial Pyramid Pooling
• Better ImageNet pretrain weights

Model Zoo

PP-YOLO

Notes:

• PP-YOLO is trained on COCO train2017 dataset and evaluated on val2017 & test-dev2017 dataset，Box AP^test is evaluation results of mAP(IoU=0.5:0.95).
• PP-YOLO used 8 GPUs for training and mini-batch size as 24 on each GPU, if GPU number and mini-batch size is changed, learning rate and iteration times should be adjusted according FAQ.
• PP-YOLO inference speed is tesed on single Tesla V100 with batch size as 1, CUDA 10.2, CUDNN 7.5.1, TensorRT 5.1.2.2 in TensorRT mode.
• PP-YOLO FP32 inference speed testing uses inference model exported by tools/export_model.py and benchmarked by running depoly/python/infer.py with --run_benchmark. All testing results do not contains the time cost of data reading and post-processing(NMS), which is same as YOLOv4(AlexyAB) in testing method.
• TensorRT FP16 inference speed testing exclude the time cost of bounding-box decoding(yolo_box) part comparing with FP32 testing above, which means that data reading, bounding-box decoding and post-processing(NMS) is excluded(test method same as YOLOv4(AlexyAB) too)

PP-YOLO for mobile

Notes:

• PP-YOLO_MobileNetV3 is trained on COCO train2017 datast and evaluated on val2017 dataset，Box AP^val is evaluation results of mAP(IoU=0.5:0.95), Box AP50^val is evaluation results of mAP(IoU=0.5).
• PP-YOLO_MobileNetV3 used 4 GPUs for training and mini-batch size as 32 on each GPU, if GPU number and mini-batch size is changed, learning rate and iteration times should be adjusted according FAQ.
• PP-YOLO_MobileNetV3 inference speed is tested on Kirin 990 with 1 thread.

PP-YOLO tiny

Notes:

• PP-YOLO-tiny is trained on COCO train2017 datast and evaluated on val2017 dataset，Box AP^val is evaluation results of mAP(IoU=0.5:0.95).
• PP-YOLO-tiny used 8 GPUs for training and mini-batch size as 32 on each GPU, if GPU number and mini-batch size is changed, learning rate and iteration times should be adjusted according FAQ.
• PP-YOLO-tiny inference speed is tested on Kirin 990 with 4 threads by arm8
• we alse provide PP-YOLO-tiny post quant inference model, which can compress model to 1.3MB with nearly no inference on inference speed and performance

PP-YOLO on Pascal VOC

PP-YOLO trained on Pascal VOC dataset as follows:

Getting Start

1. Training

Training PP-YOLO on 8 GPUs with following command(all commands should be run under PaddleDetection dygraph directory as default)

python -m paddle.distributed.launch --log_dir=./ppyolo_dygraph/ --gpus 0,1,2,3,4,5,6,7 tools/train.py -c configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml &>ppyolo_dygraph.log 2>&1 &

optional: Run tools/anchor_cluster.py to get anchors suitable for your dataset, and modify the anchor setting in model configuration file and reader configuration file, such as configs/ppyolo/_base_/ppyolo_tiny.yml and configs/ppyolo/_base_/ppyolo_tiny_reader.yml.

python tools/anchor_cluster.py -c configs/ppyolo/ppyolo_tiny_650e_coco.yml -n 9 -s 320 -m v2 -i 1000

2. Evaluation

Evaluating PP-YOLO on COCO val2017 dataset in single GPU with following commands:

# use weights released in PaddleDetection model zoo
CUDA_VISIBLE_DEVICES=0 python tools/eval.py -c configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyolo_r50vd_dcn_1x_coco.pdparams

# use saved checkpoint in training
CUDA_VISIBLE_DEVICES=0 python tools/eval.py -c configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml -o weights=output/ppyolo_r50vd_dcn_1x_coco/model_final

For evaluation on COCO test-dev2017 dataset, configs/ppyolo/ppyolo_test.yml should be used, please download COCO test-dev2017 dataset from COCO dataset download and decompress to pathes configured by EvalReader.dataset in configs/ppyolo/ppyolo_test.yml and run evaluation by following command:

# use weights released in PaddleDetection model zoo
CUDA_VISIBLE_DEVICES=0 python tools/eval.py -c configs/ppyolo/ppyolo_test.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyolo_r50vd_dcn_1x_coco.pdparams

# use saved checkpoint in training
CUDA_VISIBLE_DEVICES=0 python tools/eval.py -c configs/ppyolo/ppyolo_test.yml -o weights=output/ppyolo_r50vd_dcn_1x_coco/model_final

Evaluation results will be saved in bbox.json, compress it into a zip package and upload to COCO dataset evaluation to evaluate.

NOTE 1: configs/ppyolo/ppyolo_test.yml is only used for evaluation on COCO test-dev2017 dataset, could not be used for training or COCO val2017 dataset evaluating.

NOTE 2: Due to the overall upgrade of the dynamic graph framework, the following weight models published by paddledetection need to be evaluated by adding the -- bias field, such as

# use weights released in PaddleDetection model zoo
CUDA_VISIBLE_DEVICES=0 python tools/eval.py -c configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyolo_r50vd_dcn_1x_coco.pdparams --bias

These models are:

1.ppyolo_r50vd_dcn_1x_coco

2.ppyolo_r50vd_dcn_voc

3.ppyolo_r18vd_coco

4.ppyolo_mbv3_large_coco

5.ppyolo_mbv3_small_coco

6.ppyolo_tiny_650e_coco

3. Inference

Inference images in single GPU with following commands, use --infer_img to inference a single image and --infer_dir to inference all images in the directory.

# inference single image
CUDA_VISIBLE_DEVICES=0 python tools/infer.py -c configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyolo_r50vd_dcn_1x_coco.pdparams --infer_img=demo/000000014439_640x640.jpg

# inference all images in the directory
CUDA_VISIBLE_DEVICES=0 python tools/infer.py -c configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyolo_r50vd_dcn_1x_coco.pdparams --infer_dir=demo

4. Inferece deployment

For inference deployment or benchmard, model exported with tools/export_model.py should be used and perform inference with Paddle inference library with following commands:

# export model, model will be save in output/ppyolo as default
python tools/export_model.py -c configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyolo_r50vd_dcn_1x_coco.pdparams

# inference with Paddle Inference library
CUDA_VISIBLE_DEVICES=0 python deploy/python/infer.py --model_dir=output_inference/ppyolo_r50vd_dcn_1x_coco --image_file=demo/000000014439_640x640.jpg --use_gpu=True

Future work

1. more PP-YOLO tiny model
2. PP-YOLO model with more backbones

Appendix

Optimizing method and ablation experiments of PP-YOLO compared with YOLOv3.

Notes:

• Performance and inference spedd are measure with input shape as 608
• All models are trained on COCO train2017 datast and evaluated on val2017 & test-dev2017 dataset，Box AP is evaluation results as mAP(IoU=0.5:0.95).
• Inference speed is tested on single Tesla V100 with batch size as 1 following test method and environment configuration in benchmark above.
• YOLOv3-DarkNet53 with mAP as 39.0 is optimized YOLOv3 model in PaddleDetection，see Model Zoo for details.

Citation

@misc{long2020ppyolo,
title={PP-YOLO: An Effective and Efficient Implementation of Object Detector},
author={Xiang Long and Kaipeng Deng and Guanzhong Wang and Yang Zhang and Qingqing Dang and Yuan Gao and Hui Shen and Jianguo Ren and Shumin Han and Errui Ding and Shilei Wen},
year={2020},
eprint={2007.12099},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
@misc{ppdet2019,
title={PaddleDetection, Object detection and instance segmentation toolkit based on PaddlePaddle.},
author={PaddlePaddle Authors},
howpublished = {\url{https://github.com/PaddlePaddle/PaddleDetection}},
year={2019}
}