库和模型

日志

• 添加工厂类注册，来源cosmos
• 添加人群密度计数DM-Count
• 添加LPR，车牌检测网络retinaLP检测框和关键点模型来源，车牌识别网络LPRNet
• 添加nanodet多种backbone，EfficientNet-lite Rep-VGG，dynamic batch不支持
• 添加原工程的centernet、dbface，dynamic batch 未支持
• 添加原工程插件，解析不用Tensorrt中nvonnxparser.lib，加入onnx-tensorrt便于插件修改
• 添加人脸相似度计算，参考mnn_example整合人脸代码，检测-对齐-关键点-性别|年龄-相似度
• 添加landmark、gender-age
• 添加ghostnet
• 添加arcface
• 添加retinaface，使用pytorch版，mxnet版
• 添加centerface
• 添加nanodet，dynamic batch 未支持
• 更新库lean及模型（cuda11.0、cudnn8.0.5、tensorRT7.2.2.3，转onnx流程）
• YOLOv3-SPP、YOLOv4检测流程同YOLOv5，仅修改config即可
• 添加yaml读取配置文件
• 添加YOLOv5，直接使用版focus未优化
• onnx的plugin暂时移除，cmake编译未修改

=======================================================================

=======================================================================

=======================================================================

TensorRT-Integrate

1. Support pytorch onnx plugin（DCN、HSwish ... etc.）
2. Simpler inference and plugin APIs

Re-implement

CenterNet : ctdet_coco_dla_2x

CenterTrack: coco_tracking

• coco_tracking.onnx download

• nuScenes_3Dtracking.onnx download

DBFace

Use TensorRT-Integrate

install protobuf == 3.11.4 (or >= 3.8.x, But it's more troublesome)

bash scripts/getALL.sh
make run -j32

Inference Code

auto engine = TRTInfer::loadEngine("models/efficientnet-b0.fp32.trtmodel");
float mean[3] = {0.485, 0.456, 0.406};
float std[3] = {0.229, 0.224, 0.225};
Mat image = imread("img.jpg");
auto input = engine->input();

// multi batch sample
input->resize(2);
input->setNormMatGPU(0, image, mean, std);
input->setNormMatGPU(1, image, mean, std);

engine->forward();

// get result and copy to cpu
engine->output(0)->cpu<float>();
engine->tensor("hm")->cpu<float>();

Environment

• tensorRT7.0 or tensorRT6.0
• opencv3.4.6
• cudnn7.6.3
• cuda10.0
• protobuf v3.8.x
• Visual Studio 2017
• lean-windows.zip (include tensorRT、opencv、cudnn、cuda、protobuf)

Plugin

1. Pytorch export ONNX: plugin_onnx_export.py
2. MReLU.cu 、HSwish.cu、DCNv2.cu