English | [简体中文](README_cn.md)

# PP-YOLOE

## Table of Contents
- [Introduction](#Introduction)
- [Model Zoo](#Model-Zoo)
- [Getting Start](#Getting-Start)
- [Appendix](#Appendix)

## Introduction
PP-YOLOE is an excellent single-stage anchor-free model based on PP-YOLOv2, surpassing a variety of popular yolo models. PP-YOLOE has a series of models, named s/m/l/x, which are configured through width multiplier and depth multiplier. PP-YOLOE avoids using special operators, such as deformable convolution or matrix nms, to be deployed friendly on various hardware. For more details, please refer to our [report](https://arxiv.org/abs/2203.16250).

<div align="center">
  <img src="../../docs/images/ppyoloe_map_fps.png" width=500 />
</div>

PP-YOLOE-l achieves 51.4 mAP on COCO test-dev2017 dataset with 78.1 FPS on Tesla V100. While using TensorRT FP16, PP-YOLOE-l can be further accelerated to 149.2 FPS. PP-YOLOE-s/m/x also have excellent accuracy and speed performance, which can be found in [Model Zoo](#Model-Zoo)

PP-YOLOE is composed of following methods:
- Scalable backbone and neck
- [Task Alignment Learning](https://arxiv.org/abs/2108.07755)
- Efficient Task-aligned head with [DFL](https://arxiv.org/abs/2006.04388) and [VFL](https://arxiv.org/abs/2008.13367)
- [SiLU activation function](https://arxiv.org/abs/1710.05941)

## Model Zoo
|          Model           | GPU number | images/GPU |  backbone  | input shape | Box AP<sup>val</sup> | Box AP<sup>test</sup> | Params(M) | FLOPs(G) | V100 FP32(FPS) | V100 TensorRT FP16(FPS) | download | config  |
|:------------------------:|:-------:|:----------:|:----------:| :-------:| :------------------: | :-------------------: |:---------:|:--------:| :------------: | :---------------------: | :------: | :------: |
| PP-YOLOE-s                  |     8      |     32     | cspresnet-s |     640     |       42.7        |        43.1         |   7.93    |  17.36   |      208.3      |          333.3          | [model](https://paddledet.bj.bcebos.com/models/ppyoloe_crn_s_300e_coco.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyoloe/ppyoloe_crn_s_300e_coco.yml)                   |
| PP-YOLOE-m                  |     8      |     28     | cspresnet-m |     640     |       48.6        |        48.9         |   23.43   |  49.91   |      123.4      |          208.3          | [model](https://paddledet.bj.bcebos.com/models/ppyoloe_crn_m_300e_coco.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyoloe/ppyoloe_crn_m_300e_coco.yml)                   |
| PP-YOLOE-l                  |     8      |     20      | cspresnet-l |     640     |       50.9        |        51.4         |   52.20   |  110.07  |      78.1      |          149.2          | [model](https://paddledet.bj.bcebos.com/models/ppyoloe_crn_l_300e_coco.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyoloe/ppyoloe_crn_l_300e_coco.yml)                   |
| PP-YOLOE-x                  |     8      |     16     | cspresnet-x |     640     |       51.9        |        52.2         |   98.42   |  206.59  |      45.0      |          95.2          | [model](https://paddledet.bj.bcebos.com/models/ppyoloe_crn_x_300e_coco.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyoloe/ppyoloe_crn_x_300e_coco.yml)                   |

**Notes:**

- PP-YOLOE is trained on COCO train2017 dataset and evaluated on val2017 & test-dev2017 dataset，Box AP<sup>test</sup> is evaluation results of `mAP(IoU=0.5:0.95)`.
- PP-YOLOE used 8 GPUs for mixed precision training, if **GPU number** or **mini-batch size** is changed, **learning rate** should be adjusted according to the formula **lr<sub>new</sub> = lr<sub>default</sub> * (batch_size<sub>new</sub> * GPU_number<sub>new</sub>) / (batch_size<sub>default</sub> * GPU_number<sub>default</sub>)**.
- PP-YOLOE inference speed is tesed on single Tesla V100 with batch size as 1, **CUDA 10.2**, **CUDNN 7.6.5**, **TensorRT 6.0.1.8** in TensorRT mode.
- Refer to [Speed testing](#Speed-testing) to reproduce the speed testing results of PP-YOLOE.
- If you set `--run_benchmark=True`，you should install these dependencies at first, `pip install pynvml psutil GPUtil`.

## Getting Start

### Training

Training PP-YOLOE with mixed precision on 8 GPUs with following command

```bash
python -m paddle.distributed.launch --gpus 0,1,2,3,4,5,6,7 tools/train.py -c configs/ppyoloe/ppyoloe_crn_l_300e_coco.yml --amp
```

**Notes:**
- use `--amp` to train with default config to avoid out of memeory.
- PaddleDetection supports multi-machine distribued training, you can refer to [DistributedTraining tutorial](../../docs/DistributedTraining_en.md).


### Evaluation

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

```bash
CUDA_VISIBLE_DEVICES=0 python tools/eval.py -c configs/ppyoloe/ppyoloe_crn_l_300e_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_crn_l_300e_coco.pdparams
```

For evaluation on COCO test-dev2017 dataset, please download COCO test-dev2017 dataset from [COCO dataset download](https://cocodataset.org/#download) and decompress to COCO dataset directory and configure `EvalDataset` like `configs/ppyolo/ppyolo_test.yml`.

### 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.

```bash
# inference single image
CUDA_VISIBLE_DEVICES=0 python tools/infer.py -c configs/ppyoloe/ppyoloe_crn_l_300e_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_crn_l_300e_coco.pdparams --infer_img=demo/000000014439_640x640.jpg

# inference all images in the directory
CUDA_VISIBLE_DEVICES=0 python tools/infer.py -c configs/ppyoloe/ppyoloe_crn_l_300e_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_crn_l_300e_coco.pdparams --infer_dir=demo
```

### Exporting models

For deployment on GPU or speed testing, model should be first exported to inference model using `tools/export_model.py`.

**Exporting PP-YOLOE for Paddle Inference without TensorRT**, use following command

```bash
python tools/export_model.py -c configs/ppyoloe/ppyoloe_crn_l_300e_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_crn_l_300e_coco.pdparams
```

**Exporting PP-YOLOE for Paddle Inference with TensorRT** for better performance, use following command with extra `-o trt=True` setting.

```bash
python tools/export_model.py -c configs/ppyoloe/ppyoloe_crn_l_300e_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_crn_l_300e_coco.pdparams trt=True
```

If you want to export PP-YOLOE model to **ONNX format**, use following command refer to [PaddleDetection Model Export as ONNX Format Tutorial](../../deploy/EXPORT_ONNX_MODEL_en.md).

```bash
# export inference model
python tools/export_model.py -c configs/ppyoloe/ppyoloe_crn_l_300e_coco.yml --output_dir=output_inference -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_crn_l_300e_coco.pdparams

# install paddle2onnx
pip install paddle2onnx

# convert to onnx
paddle2onnx --model_dir output_inference/ppyoloe_crn_l_300e_coco --model_filename model.pdmodel --params_filename model.pdiparams --opset_version 11 --save_file ppyoloe_crn_l_300e_coco.onnx

```

**Notes:** ONNX model only supports batch_size=1 now

### Speed testing

For fair comparison, the speed in [Model Zoo](#Model-Zoo) do not contains the time cost of data reading and post-processing(NMS), which is same as [YOLOv4(AlexyAB)](https://github.com/AlexeyAB/darknet) in testing method. Thus, you should export model with extra `-o exclude_nms=True` setting.

**Using Paddle Inference without TensorRT** to test speed, run following command

```bash
# export inference model
python tools/export_model.py -c configs/ppyoloe/ppyoloe_crn_l_300e_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_crn_l_300e_coco.pdparams exclude_nms=True

# speed testing with run_benchmark=True
CUDA_VISIBLE_DEVICES=0 python deploy/python/infer.py --model_dir=output_inference/ppyoloe_crn_l_300e_coco --image_file=demo/000000014439_640x640.jpg --run_mode=paddle --device=gpu --run_benchmark=True
```

**Using Paddle Inference with TensorRT** to test speed, run following command

```bash
# export inference model with trt=True
python tools/export_model.py -c configs/ppyoloe/ppyoloe_crn_l_300e_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_crn_l_300e_coco.pdparams exclude_nms=True trt=True

# speed testing with run_benchmark=True,run_mode=trt_fp32/trt_fp16
CUDA_VISIBLE_DEVICES=0 python deploy/python/infer.py --model_dir=output_inference/ppyoloe_crn_l_300e_coco --image_file=demo/000000014439_640x640.jpg --run_mode=trt_fp16 --device=gpu --run_benchmark=True

```

**Using TensorRT Inference with ONNX** to test speed, run following command

```bash
# export inference model with trt=True
python tools/export_model.py -c configs/ppyoloe/ppyoloe_crn_s_300e_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_crn_s_300e_coco.pdparams exclude_nms=True trt=True

# convert to onnx
paddle2onnx --model_dir output_inference/ppyoloe_crn_s_300e_coco --model_filename model.pdmodel --params_filename model.pdiparams --opset_version 12 --save_file ppyoloe_crn_s_300e_coco.onnx

# trt inference using fp16 and batch_size=1
trtexec --onnx=./ppyoloe_crn_s_300e_coco.onnx --saveEngine=./ppyoloe_s_bs1.engine --workspace=1024 --avgRuns=1000 --shapes=image:1x3x640x640,scale_factor:1x2 --fp16

# trt inference using fp16 and batch_size=32
trtexec --onnx=./ppyoloe_crn_s_300e_coco.onnx --saveEngine=./ppyoloe_s_bs32.engine --workspace=1024 --avgRuns=1000 --shapes=image:32x3x640x640,scale_factor:32x2 --fp16

# Using the above script, T4 and tensorrt 7.2 machine, the speed of PPYOLOE-s model is as follows,

# batch_size=1, 2.80ms, 357fps
# batch_size=32, 67.69ms, 472fps

```


### Deployment

PP-YOLOE can be deployed by following approches:
  - Paddle Inference [Python](../../deploy/python) & [C++](../../deploy/cpp)
  - [Paddle-TensorRT](../../deploy/TENSOR_RT.md)
  - [PaddleServing](https://github.com/PaddlePaddle/Serving)
  - [PaddleSlim](../slim)

Next, we will introduce how to use Paddle Inference to deploy PP-YOLOE models in TensorRT FP16 mode.

First, refer to [Paddle Inference Docs](https://www.paddlepaddle.org.cn/inference/master/user_guides/download_lib.html#python), download and install packages corresponding to CUDA, CUDNN and TensorRT version.

Then, Exporting PP-YOLOE for Paddle Inference **with TensorRT**, use following command.

```bash
python tools/export_model.py -c configs/ppyoloe/ppyoloe_crn_l_300e_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_crn_l_300e_coco.pdparams trt=True
```

Finally, inference in TensorRT FP16 mode.

```bash
# inference single image
CUDA_VISIBLE_DEVICES=0 python deploy/python/infer.py --model_dir=output_inference/ppyoloe_crn_l_300e_coco --image_file=demo/000000014439_640x640.jpg --device=gpu --run_mode=trt_fp16

# inference all images in the directory
CUDA_VISIBLE_DEVICES=0 python deploy/python/infer.py --model_dir=output_inference/ppyoloe_crn_l_300e_coco --image_dir=demo/ --device=gpu  --run_mode=trt_fp16

```

**Notes:**
- TensorRT will perform optimization for the current hardware platform according to the definition of the network, generate an inference engine and serialize it into a file. This inference engine is only applicable to the current hardware hardware platform. If your hardware and software platform has not changed, you can set `use_static=True` in [enable_tensorrt_engine](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.4/deploy/python/infer.py#L660). In this way, the serialized file generated will be saved in the `output_inference` folder, and the saved serialized file will be loaded the next time when TensorRT is executed.
- PaddleDetection release/2.4 and later versions will support NMS calling TensorRT, which requires PaddlePaddle release/2.3 and later versions.

### Other Datasets

Model | AP | AP<sub>50</sub>
---|---|---
[YOLOX](https://github.com/Megvii-BaseDetection/YOLOX) | 22.6 | 37.5
[YOLOv5](https://github.com/ultralytics/yolov5) | 26.0 | 42.7
**PP-YOLOE** | **30.5** | **46.4**

**Notes**
- Here, we use [VisDrone](https://github.com/VisDrone/VisDrone-Dataset) dataset, and to detect 9 objects including `person, bicycles, car, van, truck, tricyle, awning-tricyle, bus, motor`.
- Above models trained using official default config, and load pretrained parameters on COCO dataset.
- *Due to the limited time, more verification results will be supplemented in the future. You are also welcome to contribute to PP-YOLOE*


### Feature Models

The PaddleDetection team provides configs and weights of various feature detection models based on PP-YOLOE, which users can download for use:

|Scenarios | Related Datasets | Links|
| :--------: | :---------: | :------: |
|Pedestrian Detection | CrowdHuman | [pphuman](../pphuman) |
|Vehicle Detection | BDD100K，UA-DETRAC | [ppvehicle](../ppvehicle) |
|Small Object Detection | VisDrone | [visdrone](../visdrone) |


## Appendix

Ablation experiments of PP-YOLOE.

| NO.  |        Model                 | Box AP<sup>val</sup> | Params(M) | FLOPs(G) | V100 FP32 FPS |
| :--: | :---------------------------: | :------------------: | :-------: | :------: | :-----------: |
|  A   | PP-YOLOv2          |         49.1         |   54.58   |  115.77   |     68.9     |
|  B   | A + Anchor-free    |         48.8         |   54.27   |  114.78   |     69.8     |
|  C   | B + CSPRepResNet   |         49.5         |   47.42   |  101.87   |     85.5     |
|  D   | C + TAL            |         50.4         |   48.32   |  104.75   |     84.0     |
|  E   | D + ET-Head        |         50.9         |   52.20   |  110.07   |     78.1     |