cuda-driver-api-1.1-cuinit
cuda-driver-api-1.2-check1
cuda-driver-api-1.3-check2
cuda-driver-api-1.4-context
cuda-driver-api-1.5-memory-alloc
cuda-runtime-api-1.1-hello-runtime
cuda-runtime-api-1.10-warpaffine
cuda-runtime-api-1.11-cublas-gemm
cuda-runtime-api-1.12-yolov5-postprocess
cuda-runtime-api-1.13-thrust
cuda-runtime-api-1.14-error
cuda-runtime-api-1.15-bank-conflict
cuda-runtime-api-1.2-memory
cuda-runtime-api-1.3-stream
cuda-runtime-api-1.4-kernel-function
cuda-runtime-api-1.5-thread-layout
cuda-runtime-api-1.5.2-parallel
cuda-runtime-api-1.6-vector-add
cuda-runtime-api-1.7-shared-memory
cuda-runtime-api-1.8-reduce-sum
cuda-runtime-api-1.9-atomic
tensorrt-basic-1.1-hello-tensorrt
tensorrt-basic-1.10-3rd-plugin
tensorrt-basic-1.2-hello-inference
tensorrt-basic-1.3-cnn-and-dynamic-shape
tensorrt-basic-1.4-onnx-editor
tensorrt-basic-1.5-onnx-parser
tensorrt-basic-1.6-onnx-parser-source-code
tensorrt-basic-1.7-hello-plugin
tensorrt-basic-1.8-integrate-easyplugin
tensorrt-integrate-1.1-full-cnn-classifier
tensorrt-integrate-1.10-yolov5-obb
tensorrt-integrate-1.11-onnxruntime
tensorrt-integrate-1.12-multithread
tensorrt-integrate-1.13-builder
tensorrt-integrate-1.14-memory
tensorrt-integrate-1.15-tensor
tensorrt-integrate-1.16-infer
tensorrt-integrate-1.17-multi-thread-yolov5
tensorrt-integrate-1.18-integrate-full-yolov5
tensorrt-integrate-1.19-insightface
tensorrt-integrate-1.2-yolov5-detect
tensorrt-integrate-1.20-self-driving
tensorrt-integrate-1.21-multi-camera-decoder
tensorrt-integrate-1.22-resful-http
tensorrt-integrate-1.23-openvino-yolov5
tensorrt-integrate-1.24-rknn
tensorrt-integrate-1.25-pybind11
tensorrt-integrate-1.26-openvino-integrate
tensorrt-integrate-1.27-lua
tensorrt-integrate-1.3-yolox-detect
tensorrt-integrate-1.4-retinaface-detect
tensorrt-integrate-1.5-unet
tensorrt-integrate-1.6-chinese-classifer-bert
tensorrt-integrate-1.7-huggingface-ner
tensorrt-integrate-1.8-alphapose
tensorrt-integrate-1.9-mmdetection-yolox
Latest commit dujw
Jul 24, 2022 This branch is up to date with jinmin527/learning-cuda-trt:main .
Folders and files Name Name Last commit message
Last commit date
parent directory Jul 24, 2022
Jul 24, 2022
Jul 24, 2022
Jul 24, 2022
Jul 24, 2022
Jul 24, 2022
Jul 24, 2022
Jul 24, 2022
Jul 24, 2022
View all files
yolox的预处理部分,使用了仿射变换,请参照仿射变换原理
使用仿射变换实现letterbox的理由是
便于操作,得到变换矩阵即可
便于逆操作,实则是逆矩阵映射即可
便于cuda加速,cuda版本的加速已经在cuda系列中提到了warpaffine实现
该加速可以允许warpaffine、normalize、除以255、减均值除以标准差、变换RB通道等等在一个核中实现,性能最好
后处理部分,反算到图像坐标,实际上是乘以逆矩阵
而由于逆矩阵实际上有效自由度是3,也就是d2i中只有3个数是不同的,其他都一样。也因此你看到的是d2i[0]、d2i[2]、d2i[5]在作用
这里通过自定义代码推理过程,后处理过程,实现模型导出
官方的onnx导出代码,要么有模型不支持,要么导出的模型是乱的,不行不行
这里提供了一个yolox的导出案例
安装mmdet环境bash install.sh
导出yolox模型bash export-yolox.sh
运行推理make run -j64
You can’t perform that action at this time.