【JetsonNano】ONNX 模型 转换 TensorRT，命令执行完了，但没有生成转换文件

1. 命令

sudo /usr/src/tensorrt/bin/trtexec --onnx=yolo.onnx --saveEngine=yolo.trt --fp16   --workspace=512 

执行输出完毕，未生成转换文件

...
[03/12/2024-15:49:21] [W] [TRT] Tactic Device request: 347MB Available: 85MB. Device memory is insufficient to use tactic.
[03/12/2024-15:49:22] [W] [TRT] Skipping tactic 3 due to insufficient memory on requested size of 347 detected for tactic 0x0000000000000004.
[03/12/2024-15:49:22] [W] [TRT] Tactic Device request: 347MB Available: 66MB. Device memory is insufficient to use tactic.
[03/12/2024-15:49:22] [W] [TRT] Skipping tactic 7 due to insufficient memory on requested size of 347 detected for tactic 0x000000000000003c.
[03/12/2024-15:49:28] [I] [TRT] Total Activation Memory: 580451328
[03/12/2024-15:49:28] [I] [TRT] Detected 1 inputs and 1 output network tensors.
Killed

2. 解决：输出问题

saveEngine 目录的问题

sudo /usr/src/tensorrt/bin/trtexec --onnx=yolo.onnx --saveEngine=/home/pi/yolo.trt --fp16   --workspace=512 

参考：https://forums.developer.nvidia.com/t/about-trtexec/165188

注意1：若编译过程中 jetson nano 内存不足导致被 kill ，建议增大交换内存
参考：https://labelnet.blog.csdn.net/article/details/136537801

注意2：模型转换时间较长，大概 20min 左右，耐心等待

3. 解决：编译内存问题

--memPoolSize=workspace:1800 指定大小，调整合适的大小进行转换，我这里是 1800 转换成功了，试了 512 和 1024 都失败了

/usr/src/tensorrt/bin/trtexec --onnx=yolo.onnx --saveEngine=yolo.trt --fp16 --plugins=./libyolo.so --memPoolSize=workspace:1800

成功示例