Could not find any similar ops to torchvision::nms. This op may not exist or may not be currently supported in TorchScript. about yolort

[mattpopovich]

The solution for this is to make sure you are including the following:

#include <torch/script.h>       // One-stop header; for torch::jit::load()
#include <torchvision/vision.h> // For torchvision NMS in model

If that doesn't work, I checked ldd and it looked like the torchvision library was not actually being included.

Failing executable:

# ldd path/to/failing/executable | grep torch 
        libc10.so => /usr/local/lib/python3.8/dist-packages/torch/lib/libc10.so (0x00007fe09b466000)
        libtorch.so => /usr/local/lib/python3.8/dist-packages/torch/lib/libtorch.so (0x00007fe089dd3000)
        libtorch_cpu.so => /usr/local/lib/python3.8/dist-packages/torch/lib/libtorch_cpu.so (0x00007fe07fe70000)
        libtorch_cuda.so => /usr/local/lib/python3.8/dist-packages/torch/lib/libtorch_cuda.so (0x00007fe030163000)
        libc10_cuda.so => /usr/local/lib/python3.8/dist-packages/torch/lib/libc10_cuda.so (0x00007fe01fe6b000)

Working executable (note libtorchvision.so):

# ldd /path/to/working/executable | grep torch
        libc10.so => /usr/local/lib/python3.8/dist-packages/torch/lib/libc10.so (0x00007f293c174000)
        libtorchvision.so => /usr/local/lib/libtorchvision.so (0x00007f293bba4000)
        libtorch.so => /usr/local/lib/python3.8/dist-packages/torch/lib/libtorch.so (0x00007f293bb9f000)
        libtorch_cpu.so => /usr/local/lib/python3.8/dist-packages/torch/lib/libtorch_cpu.so (0x00007f2931c3c000)
        libtorch_cuda.so => /usr/local/lib/python3.8/dist-packages/torch/lib/libtorch_cuda.so (0x00007f28e1f52000)
        libc10_cuda.so => /usr/local/lib/python3.8/dist-packages/torch/lib/libc10_cuda.so (0x00007f28e16d5000)

In order for libtorchvision.so to be successfully included, I had to #include <torchvision/models/resnet.h> and then actually use that in my code:

auto model = vision::models::ResNet18( );

That was the easiest way I found to actually use something from torchvision in your code... After that, the failing executable showed libtorchvision.so with ldd and I was able to successfully load my yolov5-rt-stack model with NMS post-processing in C++! 🥳

[zhiqwang]

Hi @mattpopovich , Thanks for the tip here, it was very useful. I had some similar ldd linking problems when I was first compiling torchvision.