README.md 20.2 KB
Newer Older
W
wangxinxin08 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13
English | [简体中文](README_cn.md)

# PP-YOLO

## Table of Contents
- [Introduction](#Introduction)
- [Model Zoo](#Model_Zoo)
- [Getting Start](#Getting_Start)
- [Future Work](#Future_Work)
- [Appendix](#Appendix)

## Introduction

14
[PP-YOLO](https://arxiv.org/abs/2007.12099) is a optimized model based on YOLOv3 in PaddleDetection,whose performance(mAP on COCO) and inference spped are better than [YOLOv4](https://arxiv.org/abs/2004.10934),PaddlePaddle 2.0.2(available on pip now) or [Daily Version](https://www.paddlepaddle.org.cn/documentation/docs/zh/develop/install/Tables.html#whl-develop) is required to run this PP-YOLO。
W
wangxinxin08 已提交
15 16 17 18

PP-YOLO reached mmAP(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.

<div align="center">
19
  <img src="../../docs/images/ppyolo_map_fps.png" width=500 />
W
wangxinxin08 已提交
20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39
</div>

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](https://arxiv.org/abs/1810.12890)
- [Exponential Moving Average](https://www.investopedia.com/terms/e/ema.asp)
- [IoU Loss](https://arxiv.org/pdf/1902.09630.pdf)
- [Grid Sensitive](https://arxiv.org/abs/2004.10934)
- [Matrix NMS](https://arxiv.org/pdf/2003.10152.pdf)
- [CoordConv](https://arxiv.org/abs/1807.03247)
- [Spatial Pyramid Pooling](https://arxiv.org/abs/1406.4729)
- Better ImageNet pretrain weights

## Model Zoo

### PP-YOLO

|          Model           | GPU number | images/GPU |  backbone  | input shape | Box AP<sup>val</sup> | Box AP<sup>test</sup> | V100 FP32(FPS) | V100 TensorRT FP16(FPS) | download | config  |
W
wangxinxin08 已提交
40
|:------------------------:|:-------:|:-------------:|:----------:| :-------:| :------------------: | :-------------------: | :------------: | :---------------------: | :------: | :------: |
W
wangxinxin08 已提交
41 42 43 44 45 46 47 48 49 50 51
| PP-YOLO                  |     8      |     24     | ResNet50vd |     608     |         44.8         |         45.2          |      72.9      |          155.6          | [model](https://paddledet.bj.bcebos.com/models/ppyolo_r50vd_dcn_1x_coco.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml)                   |
| PP-YOLO                  |     8      |     24     | ResNet50vd |     512     |         43.9         |         44.4          |      89.9      |          188.4          | [model](https://paddledet.bj.bcebos.com/models/ppyolo_r50vd_dcn_1x_coco.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml)                   |
| PP-YOLO                  |     8      |     24     | ResNet50vd |     416     |         42.1         |         42.5          |      109.1      |          215.4          | [model](https://paddledet.bj.bcebos.com/models/ppyolo_r50vd_dcn_1x_coco.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml)                   |
| PP-YOLO                  |     8      |     24     | ResNet50vd |     320     |         38.9         |         39.3          |      132.2      |          242.2          | [model](https://paddledet.bj.bcebos.com/models/ppyolo_r50vd_dcn_1x_coco.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml)                   |
| PP-YOLO_2x               |     8      |     24     | ResNet50vd |     608     |         45.3         |         45.9          |      72.9      |          155.6          | [model](https://paddledet.bj.bcebos.com/models/ppyolo_r50vd_dcn_2x_coco.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_2x_coco.yml)                   |
| PP-YOLO_2x               |     8      |     24     | ResNet50vd |     512     |         44.4         |         45.0          |      89.9      |          188.4          | [model](https://paddledet.bj.bcebos.com/models/ppyolo_r50vd_dcn_2x_coco.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_2x_coco.yml)                   |
| PP-YOLO_2x               |     8      |     24     | ResNet50vd |     416     |         42.7         |         43.2          |      109.1      |          215.4          | [model](https://paddledet.bj.bcebos.com/models/ppyolo_r50vd_dcn_2x_coco.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_2x_coco.yml)                   |
| PP-YOLO_2x               |     8      |     24     | ResNet50vd |     320     |         39.5         |         40.1          |      132.2      |          242.2          | [model](https://paddledet.bj.bcebos.com/models/ppyolo_r50vd_dcn_2x_coco.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_2x_coco.yml)                   |
| PP-YOLO               |     4      |     32     | ResNet18vd |     512     |         29.2         |         29.5          |      357.1      |          657.9          | [model](https://paddledet.bj.bcebos.com/models/ppyolo_r18vd_coco.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r18vd_coco.yml)                   |
| PP-YOLO               |     4      |     32     | ResNet18vd |     416     |         28.6         |         28.9          |      409.8      |          719.4          | [model](https://paddledet.bj.bcebos.com/models/ppyolo_r18vd_coco.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r18vd_coco.yml)                   |
| PP-YOLO               |     4      |     32     | ResNet18vd |     320     |         26.2         |         26.4          |      480.7      |          763.4          | [model](https://paddledet.bj.bcebos.com/models/ppyolo_r18vd_coco.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r18vd_coco.yml)                   |
52 53
| PP-YOLOv2               |     8      |     12     | ResNet50vd |     640     |         49.1         |         49.5          |      68.9      |          106.5          | [model](https://paddledet.bj.bcebos.com/models/ppyolov2_r50vd_dcn_365e_coco.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolov2_r50vd_dcn_365e_coco.yml)                   |
| PP-YOLOv2               |     8      |     12     | ResNet101vd |     640     |         49.7         |         50.3          |     49.5     |         87.0         | [model](https://paddledet.bj.bcebos.com/models/ppyolov2_r101vd_dcn_365e_coco.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolov2_r101vd_dcn_365e_coco.yml)                   |
W
wangxinxin08 已提交
54

W
wangxinxin08 已提交
55 56 57 58

**Notes:**

- PP-YOLO 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)`.
59
- 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](https://github.com/PaddlePaddle/PaddleDetection/blob/develop/docs/tutorials/FAQ.md).
W
wangxinxin08 已提交
60 61 62 63
- 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)](https://github.com/AlexeyAB/darknet) 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)](https://github.com/AlexeyAB/darknet) too)

W
wangxinxin08 已提交
64 65 66 67
### PP-YOLO for mobile

|            Model             | GPU number | images/GPU | Model Size | input shape | Box AP<sup>val</sup> |  Box AP50<sup>val</sup> | Kirin 990 1xCore(FPS) | download | config  |
|:----------------------------:|:-------:|:-------------:|:----------:| :-------:| :------------------: |  :--------------------: | :--------------------: | :------: | :------: |
W
wangxinxin08 已提交
68 69
| PP-YOLO_MobileNetV3_large    |    4    |      32       |    28MB    |   320    |         23.2         |           42.6          |           14.1         | [model](https://paddledet.bj.bcebos.com/models/ppyolo_mbv3_large_coco.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_mbv3_large_coco.yml)                   |
| PP-YOLO_MobileNetV3_small    |    4    |      32       |    16MB    |   320    |         17.2         |           33.8          |           21.5         | [model](https://paddledet.bj.bcebos.com/models/ppyolo_mbv3_small_coco.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_mbv3_small_coco.yml)                   |
W
wangxinxin08 已提交
70 71 72 73

**Notes:**

- PP-YOLO_MobileNetV3 is trained on COCO train2017 datast and evaluated on val2017 dataset,Box AP<sup>val</sup> is evaluation results of `mAP(IoU=0.5:0.95)`, Box AP<sup>val</sup> is evaluation results of `mAP(IoU=0.5)`.
74
- 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](https://github.com/PaddlePaddle/PaddleDetection/blob/develop/docs/tutorials/FAQ.md).
W
wangxinxin08 已提交
75 76
- PP-YOLO_MobileNetV3 inference speed is tested on Kirin 990 with 1 thread.

K
Kaipeng Deng 已提交
77 78 79 80 81 82 83 84 85 86
### PP-YOLO tiny

|            Model             | GPU number | images/GPU | Model Size | Post Quant Model Size | input shape | Box AP<sup>val</sup> | Kirin 990 4xCore(FPS) | download | config | post quant model |
|:----------------------------:|:-------:|:-------------:|:----------:| :-------------------: | :---------: | :------------------: | :-------------------: | :------: | :----: | :--------------: |
| PP-YOLO tiny                 |    8    |      32       |   4.2MB    |       **1.3M**        |     320     |         20.6         |          92.3         | [model](https://paddledet.bj.bcebos.com/models/ppyolo_tiny_650e_coco.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_tiny_650e_coco.yml)  | [inference model](https://paddledet.bj.bcebos.com/models/ppyolo_tiny_quant.tar) |
| PP-YOLO tiny                 |    8    |      32       |   4.2MB    |       **1.3M**        |     416     |         22.7         |          65.4         | [model](https://paddledet.bj.bcebos.com/models/ppyolo_tiny_650e_coco.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_tiny_650e_coco.yml)  | [inference model](https://paddledet.bj.bcebos.com/models/ppyolo_tiny_quant.tar) |

**Notes:**

- PP-YOLO-tiny is trained on COCO train2017 datast and evaluated on val2017 dataset,Box AP<sup>val</sup> is evaluation results of `mAP(IoU=0.5:0.95)`, Box AP<sup>val</sup> is evaluation results of `mAP(IoU=0.5)`.
87
- 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](https://github.com/PaddlePaddle/PaddleDetection/blob/develop/docs/tutorials/FAQ.md).
K
Kaipeng Deng 已提交
88 89 90
- 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

W
wangxinxin08 已提交
91 92 93 94 95 96
### PP-YOLO on Pascal VOC

PP-YOLO trained on Pascal VOC dataset as follows:

|       Model        | GPU number | images/GPU |  backbone  | input shape | Box AP50<sup>val</sup> | download | config  |
|:------------------:|:----------:|:----------:|:----------:| :----------:| :--------------------: | :------: | :-----: |
W
wangxinxin08 已提交
97 98 99
| PP-YOLO            |    8    |       12      | ResNet50vd |     608     |          84.9          | [model](https://paddledet.bj.bcebos.com/models/ppyolo_r50vd_dcn_voc.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_voc.yml)                   |
| PP-YOLO            |    8    |       12      | ResNet50vd |     416     |          84.3          | [model](https://paddledet.bj.bcebos.com/models/ppyolo_r50vd_dcn_voc.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_voc.yml)                   |
| PP-YOLO            |    8    |       12      | ResNet50vd |     320     |          82.2          | [model](https://paddledet.bj.bcebos.com/models/ppyolo_r50vd_dcn_voc.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_voc.yml)                   |
W
wangxinxin08 已提交
100

W
wangxinxin08 已提交
101 102 103 104 105 106 107 108 109 110
## Getting Start

### 1. Training

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

```bash
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 &
```

111 112 113 114 115 116
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`.

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

W
wangxinxin08 已提交
117 118 119 120 121 122
### 2. Evaluation

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

```bash
# use weights released in PaddleDetection model zoo
123
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
W
wangxinxin08 已提交
124 125 126 127 128 129 130 131 132

# 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](https://cocodataset.org/#download) and decompress to pathes configured by `EvalReader.dataset` in `configs/ppyolo/ppyolo_test.yml` and run evaluation by following command:

```bash
# use weights released in PaddleDetection model zoo
133
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
W
wangxinxin08 已提交
134 135 136 137 138 139 140

# 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](https://competitions.codalab.org/competitions/20794#participate) to evaluate.

141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161
**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

```bash
# 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
W
wangxinxin08 已提交
162 163 164 165 166 167 168

### 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.

```bash
# inference single image
169
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
W
wangxinxin08 已提交
170 171

# inference all images in the directory
172
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
W
wangxinxin08 已提交
173 174
```

175
### 4. Inferece deployment
W
wangxinxin08 已提交
176 177 178 179 180

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:

```bash
# export model, model will be save in output/ppyolo as default
181
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
W
wangxinxin08 已提交
182 183

# inference with Paddle Inference library
184
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
W
wangxinxin08 已提交
185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215
```


## 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.

| NO.  |        Model                 | Box AP<sup>val</sup> | Box AP<sup>test</sup> | Params(M) | FLOPs(G) | V100 FP32 FPS |
| :--: | :--------------------------- | :------------------: |:--------------------: | :-------: | :------: | :-----------: |
|  A   | YOLOv3-DarkNet53             |         38.9         |           -           |   59.13   |  65.52   |      58.2     |
|  B   | YOLOv3-ResNet50vd-DCN        |         39.1         |           -           |   43.89   |  44.71   |      79.2     |
|  C   | B + LB + EMA + DropBlock     |         41.4         |           -           |   43.89   |  44.71   |      79.2     |
|  D   | C + IoU Loss                 |         41.9         |           -           |   43.89   |  44.71   |      79.2     |
|  E   | D + IoU Aware                |         42.5         |           -           |   43.90   |  44.71   |      74.9     |
|  F   | E + Grid Sensitive           |         42.8         |           -           |   43.90   |  44.71   |      74.8     |
|  G   | F + Matrix NMS               |         43.5         |           -           |   43.90   |  44.71   |      74.8     |
|  H   | G + CoordConv                |         44.0         |           -           |   43.93   |  44.76   |      74.1     |
|  I   | H + SPP                      |         44.3         |         45.2          |   44.93   |  45.12   |      72.9     |
|  J   | I + Better ImageNet Pretrain |         44.8         |         45.2          |   44.93   |  45.12   |      72.9     |
|  K   | J + 2x Scheduler             |         45.3         |         45.9          |   44.93   |  45.12   |      72.9     |

**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.
216
- [YOLOv3-DarkNet53](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_darknet53_270e_coco.yml) with mAP as 39.0 is optimized YOLOv3 model in PaddleDetection,see [YOLOv3](https://github.com/PaddlePaddle/PaddleDetection/blob/develop/configs/yolov3/README.md) for details.
W
wangguanzhong 已提交
217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235

## 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}
}
```