- Covers [YOLO family](https://github.com/nemonameless/PaddleDetection_YOLOSeries) classic and latest models: YOLOv3, PP-YOLOE (a real-time high-precision object detection model developed by Baidu PaddlePaddle), and cutting-edge detection algorithms such as YOLOv4, YOLOv5, YOLOX, YOLOv6, and YOLOv7
- Release [PaddleYOLO](https://github.com/PaddlePaddle/PaddleYOLO) which overs classic and latest models of [YOLO family](https://github.com/PaddlePaddle/PaddleYOLO/tree/develop/docs/MODEL_ZOO_en.md): YOLOv3, PP-YOLOE (a real-time high-precision object detection model developed by Baidu PaddlePaddle), and cutting-edge detection algorithms such as YOLOv4, YOLOv5, YOLOX, YOLOv6, and YOLOv7
- Newly add high precision detection model based on [ViT](configs/vitdet) backbone network, with a 55.7% mAP accuracy on COCO dataset; newly add multi-object tracking model [OC-SORT](configs/mot/ocsort); newly add [ConvNeXt](configs/convnext) backbone network.
- All models were trained and tested in the COCO17 dataset.
- The codes of [YOLOv5](https://github.com/PaddlePaddle/PaddleYOLO/tree/develop/configs/yolov5),[YOLOv6](https://github.com/PaddlePaddle/PaddleYOLO/tree/develop/configs/yolov6) and [YOLOv7](https://github.com/PaddlePaddle/PaddleYOLO/tree/develop/configs/yolov7) can be found in [PaddleYOLO](https://github.com/PaddlePaddle/PaddleYOLO). Note that **the LICENSE of PaddleYOLO is GPL 3.0**.
- Unless special instructions, all the ResNet backbone network using [ResNet-B](https://arxiv.org/pdf/1812.01187) structure.
-**Inference time (FPS)**: The reasoning time was calculated on a Tesla V100 GPU by `tools/eval.py` testing all validation sets in FPS (number of pictures/second). CuDNN version is 7.5, including data loading, network forward execution and post-processing, and Batch size is 1.
...
...
@@ -18,132 +36,208 @@
- We adopt and [Detectron](https://github.com/facebookresearch/Detectron/blob/master/MODEL_ZOO.md#training-schedules) in the same training strategy.
- 1x strategy indicates that when the total batch size is 8, the initial learning rate is 0.01, and the learning rate decreases by 10 times after 8 epoch and 11 epoch, respectively, and the final training is 12 epoch.
- 2X strategy is twice as much as strategy 1X, and the learning rate adjustment position is twice as much as strategy 1X.
- 2x strategy is twice as much as strategy 1x, and the learning rate adjustment position of epochs is twice as much as strategy 1x.
## ImageNet pretraining model
Paddle provides a skeleton network pretraining model based on ImageNet. All pre-training models were trained by standard Imagenet 1K dataset. Res Net and Mobile Net are high-precision pre-training models obtained by cosine learning rate adjustment strategy or SSLD knowledge distillation training. Model details are available at [PaddleClas](https://github.com/PaddlePaddle/PaddleClas).
Paddle provides a skeleton network pretraining model based on ImageNet. All pre-training models were trained by standard Imagenet 1K dataset. ResNet and MobileNet are high-precision pre-training models obtained by cosine learning rate adjustment strategy or SSLD knowledge distillation training. Model details are available at [PaddleClas](https://github.com/PaddlePaddle/PaddleClas).
**PaddleYOLO** is a YOLO Series toolbox based on [PaddleDetection](https://github.com/PaddlePaddle/PaddleDetection), **only relevant codes of YOLO series models are included**. It supports `YOLOv3`,`PP-YOLO`,`PP-YOLOv2`,`PP-YOLOE`,`PP-YOLOE+`,`YOLOX`,`YOLOv5`,`YOLOv6`,`YOLOv7`,`RTMDet` and so on. Welcome to use and build it together!
## Updates
* 【2022/09/29】Support [RTMDet](https://github.com/PaddlePaddle/PaddleYOLO/tree/release/2.5/configs/rtmdet) inference and deploy;
* 【2022/09/19】Support the new version of [`YOLOv6`](https://github.com/PaddlePaddle/PaddleYOLO/tree/release/2.5/configs/yolov6), including n/t/s/m/l model;
* 【2022/08/23】Release `YOLOSeries` codebase: support `YOLOv3`,`PP-YOLOE`,`PP-YOLOE+`,`YOLOX`,`YOLOv5`,`YOLOv6` and `YOLOv7`; support using `ConvNeXt` backbone to get high-precision version of `PP-YOLOE`,`YOLOX` and `YOLOv5`; support PaddleSlim accelerated quantitative training `PP-YOLOE`,`YOLOv5`,`YOLOv6` and `YOLOv7`. For details, please read this [article](https://mp.weixin.qq.com/s/Hki01Zs2lQgvLSLWS0btrA)；
**Notes：**
- The Licence of **PaddleYOLO** is **GPL 3.0**, the codes of [YOLOv5](https://github.com/PaddlePaddle/PaddleYOLO/tree/release/2.5/configs/yolov5),[YOLOv7](https://github.com/PaddlePaddle/PaddleYOLO/tree/release/2.5/configs/yolov7) and [YOLOv6](https://github.com/PaddlePaddle/PaddleYOLO/tree/release/2.5/configs/yolov6) will not be merged into [PaddleDetection](https://github.com/PaddlePaddle/PaddleDetection). Except for these three YOLO models, other YOLO models are recommended to use in [PaddleDetection](https://github.com/PaddlePaddle/PaddleDetection), **which will be the first to release the latest progress of PP-YOLO series detection model**;
- To use **PaddleYOLO**, **PaddlePaddle-2.3.2 or above is recommended**，please refer to the [official website](https://www.paddlepaddle.org.cn/install/quick?docurl=/documentation/docs/zh/install/pip/linux-pip.html) to download the appropriate version. **For Windows platforms, please install the paddle develop version**;
- Training **Custom dataset** please refer to [doc](#CustomDataset) and [issue](https://github.com/PaddlePaddle/PaddleYOLO/issues/43). Please **ensure COCO trained weights are loaded as pre-train** at first. We recommend to use YOLO detection model **with a total `batch_size` at least greater than `64` to train**. If the resources are insufficient, please **use the smaller model** or **reduce the input size of the model**. To ensure high detection accuracy, **you'd better never try to using single GPU or total `batch_size` less than `32` for training**;
- All the models are trained on COCO train2017 dataset and evaluated on val2017 dataset. The * in front of the model indicates that the training is being updated.
- Please check the specific accuracy and speed details in [PP-YOLOE](https://github.com/PaddlePaddle/PaddleYOLO/tree/release/2.5/configs/ppyoloe),[YOLOX](https://github.com/PaddlePaddle/PaddleYOLO/tree/release/2.5/configs/yolox),[YOLOv5](https://github.com/PaddlePaddle/PaddleYOLO/tree/release/2.5/configs/yolov5),[YOLOv6](https://github.com/PaddlePaddle/PaddleYOLO/tree/release/2.5/configs/yolov6),[YOLOv7](https://github.com/PaddlePaddle/PaddleYOLO/tree/release/2.5/configs/yolov7). **Note that YOLOv5, YOLOv6 and YOLOv7 have not adopted `multi_label` to eval**.
- TRT-FP16-Latency(ms) is the time spent in testing under TensorRT-FP16, **excluding data preprocessing and model output post-processing (NMS)**. The test adopts single card **Tesla T4 GPU, batch size=1**, and the test environment is **paddlepaddle-2.3.2**, **CUDA 11.2**, **CUDNN 8.2**, **GCC-8.2**, **TensorRT 8.0.3.4**. Please refer to the respective model homepage for details.
- For **FLOPs(G) and Params(M)**, you should first install [PaddleSlim](https://github.com/PaddlePaddle/PaddleSlim), `pip install paddleslim`, then set `print_flops: True` and `print_params: True` in [runtime.yml](https://github.com/PaddlePaddle/PaddleYOLO/tree/release/2.5/configs/runtime.yml). Make sure **single scale** like 640x640, **MACs are printed，FLOPs=2*MACs**.
- Based on [PaddleSlim](https://github.com/PaddlePaddle/PaddleSlim), quantitative training of YOLO series models can achieve basically lossless accuracy and generally improve the speed by more than 30%. For details, please refer to [auto_compression](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/example/auto_compression).
- The VOC mAP is `mAP(IoU=0.5)`, and all the models **have not adopted `multi_label` to eval**.
- All YOLO VOC models are loaded with the COCO weights of their respective models as pre-train weights. Each config file uses 8 GPUs by default, which can be used as a reference for setting custom datasets. The specific mAP will vary depending on the datasets;
- We recommend to use YOLO detection model **with a total `batch_size` at least greater than `64` to train**. If the resources are insufficient, please **use the smaller model** or **reduce the input size of the model**. To ensure high detection accuracy, **you'd better not try to using single GPU or total `batch_size` less than `64` for training**;
- Params (M) and FLOPs (G) are measured during training. YOLOv7 has no s model, so tiny model is selected;
- For TRT-FP16 Latency (ms) speed measurement, please refer to the config homepage of each YOLO model;
The download link provided by PaddleDetection team is: [coco](https://bj.bcebos.com/v1/paddledet/data/coco.tar)(about 22G) and [test2017](https://bj.bcebos.com/v1/paddledet/data/cocotest2017.zip). Note that test2017 is optional, and the evaluation is based on val2017.
### **Pipeline**
Write the following commands in a script file, such as ```run.sh```, and run as：```sh run.sh```. You can also run the command line sentence by sentence.
- If you want to switch models, just modify the first two lines, such as:
```
model_name=yolov7
job_name=yolov7_tiny_300e_coco
```
- For **exporting onnx**, you should install [Paddle2ONNX](https://github.com/PaddlePaddle/Paddle2ONNX) by `pip install paddle2onnx` at first.
- For **FLOPs(G) and Params(M)**, you should install [PaddleSlim](https://github.com/PaddlePaddle/PaddleSlim) by `pip install paddleslim` at first, then set `print_flops: True` and `print_params: True` in [runtime.yml](https://github.com/PaddlePaddle/PaddleYOLO/tree/release/2.5/configs/runtime.yml). Make sure **single scale** like 640x640, **MACs are printed，FLOPs=2*MACs**.
### CustomDataset
#### preparation：
1.For the annotation of custom dataset, please refer to[DetAnnoTools](../tutorials/data/DetAnnoTools.md);
2.For training preparation of custom dataset，please refer to[PrepareDataSet](../tutorials/PrepareDataSet.md).
#### fintune：
In addition to changing the path of the dataset, it is generally recommended to load **the COCO pre training weight of the corresponding model** to fintune, which will converge faster and achieve higher accuracy, such as：
- The fintune training will show that the channels of the last layer of the head classification branch is not matched, which is a normal situation, because the number of custom dataset is generally inconsistent with that of COCO dataset;
- In general, the number of epochs for fintune training can be set less, and the lr setting is also smaller, such as 1/10. The highest accuracy may occur in one of the middle epochs;
#### Predict and export:
When using custom dataset to predict and export models, if the path of the TestDataset dataset is set incorrectly, COCO 80 categories will be used by default.
In addition to the correct path setting of the TestDataset dataset, you can also modify and add the corresponding `label_list`. Txt file (one category is recorded in one line), and `anno_path` in TestDataset can also be set as an absolute path, such as:
```
TestDataset:
!ImageFolder
anno_path: label_list.txt # if not set dataset_dir, the anno_path will be relative path of PaddleDetection root directory
# dataset_dir: dataset/my_coco # if set dataset_dir, the anno_path will be dataset_dir/anno_path
```
one line in `label_list.txt` records a corresponding category：
