Skip to content

Latest commit

 

History

History
202 lines (149 loc) · 6.72 KB

README_cn.md

File metadata and controls

202 lines (149 loc) · 6.72 KB
Raw

简体中文 | English

模型

Model Epoch backbone input shape $AP^{val}$ $AP^{val}_{50}$ Params(M) FLOPs(G) T4 TensorRT FP16(FPS) Pretrained Model config
RT-DETR-R18 6x ResNet-18 640 46.5 63.8 20 60 217 download config
RT-DETR-R34 6x ResNet-34 640 48.9 66.8 31 92 161 download config
RT-DETR-R50-m 6x ResNet-50 640 51.3 69.6 36 100 145 download config
RT-DETR-R50 6x ResNet-50 640 53.1 71.3 42 136 108 download config
RT-DETR-R101 6x ResNet-101 640 54.3 72.7 76 259 74 download config
RT-DETR-L 6x HGNetv2 640 53.0 71.6 32 110 114 download config
RT-DETR-X 6x HGNetv2 640 54.8 73.1 67 234 74 download config

注意事项:

  • RT-DETR 使用4个GPU训练。
  • RT-DETR 在COCO train2017上训练，并在val2017上评估。

快速开始

依赖包 
pip install -r requirements.txt
准备数据
训练&评估
  • 单卡GPU上训练:
# training on single-GPU
export CUDA_VISIBLE_DEVICES=0
python tools/train.py -c configs/rtdetr/rtdetr_r50vd_6x_coco.yml --eval
  • 多卡GPU上训练:
# training on multi-GPU
export CUDA_VISIBLE_DEVICES=0,1,2,3
python -m paddle.distributed.launch --gpus 0,1,2,3 tools/train.py -c configs/rtdetr/rtdetr_r50vd_6x_coco.yml --fleet --eval
  • 评估:
python tools/eval.py -c configs/rtdetr/rtdetr_r50vd_6x_coco.yml \
              -o weights=https://bj.bcebos.com/v1/paddledet/models/rtdetr_r50vd_6x_coco.pdparams
  • 测试:
python tools/infer.py -c configs/rtdetr/rtdetr_r50vd_6x_coco.yml \
              -o weights=https://bj.bcebos.com/v1/paddledet/models/rtdetr_r50vd_6x_coco.pdparams \
              --infer_img=./demo/000000570688.jpg

详情请参考快速开始文档.

部署

1. 导出模型 
python tools/export_model.py -c configs/rtdetr/rtdetr_r50vd_6x_coco.yml \
              -o weights=https://bj.bcebos.com/v1/paddledet/models/rtdetr_r50vd_6x_coco.pdparams trt=True \
              --output_dir=output_inference
2. 转换模型至ONNX 
pip install onnx==1.13.0
pip install paddle2onnx==1.0.5
  • 转换模型:
paddle2onnx --model_dir=./output_inference/rtdetr_r50vd_6x_coco/ \
            --model_filename model.pdmodel  \
            --params_filename model.pdiparams \
            --opset_version 16 \
            --save_file rtdetr_r50vd_6x_coco.onnx
3. 转换成TensorRT
  • 确保TensorRT的版本>=8.5.1
  • TRT推理可以参考RT-DETR的部分代码或者其他网络资源
trtexec --onnx=./rtdetr_r50vd_6x_coco.onnx \
        --workspace=4096 \
        --shapes=image:1x3x640x640 \
        --saveEngine=rtdetr_r50vd_6x_coco.trt \
        --avgRuns=100 \
        --fp16

其他

1. 参数量和计算量统计
  1. 找到本地安装paddle的flops源代码, 并修改为
# anaconda3/lib/python3.8/site-packages/paddle/hapi/dynamic_flops.py
def flops(net, input_size, inputs=None, custom_ops=None, print_detail=False):
    if isinstance(net, nn.Layer):
        # If net is a dy2stat model, net.forward is StaticFunction instance,
        # we set net.forward to original forward function.
        _, net.forward = unwrap_decorators(net.forward)

        # by lyuwenyu
        if inputs is None:
            inputs = paddle.randn(input_size)

        return dynamic_flops(
            net, inputs=inputs, custom_ops=custom_ops, print_detail=print_detail
        )
    elif isinstance(net, paddle.static.Program):
        return static_flops(net, print_detail=print_detail)
    else:
        warnings.warn(
            "Your model must be an instance of paddle.nn.Layer or paddle.static.Program."
        )
        return -1
  1. 使用以下代码片段实现参数量和计算量的统计
import paddle
from ppdet.core.workspace import load_config, merge_config
from ppdet.core.workspace import create

cfg_path = './configs/rtdetr/rtdetr_r50vd_6x_coco.yml'
cfg = load_config(cfg_path)
model = create(cfg.architecture)

blob = {
    'image': paddle.randn([1, 3, 640, 640]),
    'im_shape': paddle.to_tensor([[640, 640]]),
    'scale_factor': paddle.to_tensor([[1., 1.]])
}
paddle.flops(model, None, blob, custom_ops=None, print_detail=False)
2. YOLOs端到端速度测速
  • 可以参考RT-DETR benchmark部分或者其他网络资源

引用RT-DETR

如果需要在你的研究中使用RT-DETR，请通过以下方式引用我们的论文：

@misc{lv2023detrs,
      title={DETRs Beat YOLOs on Real-time Object Detection},
      author={Wenyu Lv and Shangliang Xu and Yian Zhao and Guanzhong Wang and Jinman Wei and Cheng Cui and Yuning Du and Qingqing Dang and Yi Liu},
      year={2023},
      eprint={2304.08069},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}