Unable to run evaluation on ONNX model due to unregistered MMCVDeformConv2d
#439
Assignees
Comments
|
There should be a from mmdeploy.backend.onnxruntime import get_ops_path
get_ops_path()If there is no such library. You might miss something when you install the repo. Such as : wget https://github.com/microsoft/onnxruntime/releases/download/v1.8.1/onnxruntime-linux-x64-1.8.1.tgz
tar -zxvf onnxruntime-linux-x64-1.8.1.tgz
cd onnxruntime-linux-x64-1.8.1
export ONNXRUNTIME_DIR=$(pwd)
export LD_LIBRARY_PATH=$ONNXRUNTIME_DIR/lib:$LD_LIBRARY_PATHOr cmake ...... MMDEPLOY_TARGET_BACKENDS="ort" ..... |
|
Hey @grimoire Thanks for your quick reply. I do find the
If clue for the following warning when testing the ONNX model? |
|
In fact, I just tried building with the official Docker file here. I ran conversion and evaluation inside the docker, and I had the same error message: |
|
Oh, sorry we do not have an implementation of ONNX Runtime |
|
Thanks for your clarification, @grimoire . I can try the implementation in MMCV as you suggested. The TensorRT engine would also work for us. That was actually the first engine that I tried. When I tried that, I encountered the following error during onnx2tensorrt. That's why I switched to the onnx engine. Do you know what would be the potential root cause? Here's the full log: |
|
Could you provide more detail about this conversion? |
|
@grimoire Sure, I trained a RepPoints model with ResNet50 backbone through MMDetection. I was trying to convert it to PyTorch model in a NVIDIA V100 GPU for deployment. I did convertion by following the example here: And encountered an error message: My hardware and environment info is reported in the issue description. Would there be any other info that you need? |
|
I have just add support to reppoints in #457. |
|
I just tested your PR in my machine and it worked perfectly :) The conversion was successful and the final metric looks good! Thanks, @grimoire One quick question: is there any suggestion/guiding material on how to pick the optimal |
|
Input shape should be in the range of the given config (min< [your tensor size] < max). It is always a good idea to use a static config since optimization of static shape is better. |
|
@Ivan-Zhou Hi, |
I met the same issue too. Any solution? Thanks. |
I faced the same issue, Could anyone please share the solution? |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
My goal is to convert a Torch model trained via mmdetection to ONNX. I first followed the guide to build toolchains, dependencies, and MMDeploy in my docker environment. Following this, I've successfully convert from a Torch model to an ONNX model through
deploy/tools/test.py. Then when I ran evaluation by following this page, it throws an error that the operationMMCVDeformConv2dis not registered.According to here, my understanding is that
MMCVModulatedDeformConv2dis supported as a custom ONNX operation in MMDeploy. So I want to post the issue here to see if there's any gap in my implementation with the instruction.Here's the script I ran to convert torch2onnx:
Here's the script I ran to evaluate the onnx model:
Here's my environment
Log of converting to ONNX:
Log of evaluating the ONNX model:
The text was updated successfully, but these errors were encountered: