How to export a detectron2 model into onnx format ?

[ghost]

Ask a Question

Question

I'm trying to export a detectron2 model in onnx format (model.onnx) and do inference with onnxruntime using the exported file (model.onnx).

To do so, i tried to export the famous instance segmentation model provided by detectron2 ( model_zoo ), i succeeded to get the model.onnx file ( i'm not sure if i did correctly though ) but i can't use it to do inference, can anyone tell me what i'm doing wrong ?

Here is a shared google colab notebook containing my code https://colab.research.google.com/drive/11BB-iufQCmwHoRzS4iBFwH7SqasDOik8?usp=sharing

Note 1 : In order to modify the file, you'll have to make a copy of it in your drive !
Note 2 : I raised this issue in the onnx github page, and their answer was :

Hi @mountassir-el,
As the error mentioned, AliasWithName is not an official ONNX op so there is no registered schema for it.
You can check all official ONNX ops here: https://github.com/onnx/onnx/blob/master/docs/Operators.md Thus,
this should be a question for the converter.
Please raise this question in detectron2 to ask the exporter only to produce official ONNX ops. 
Close here now. Please reopen it if you still have other ONNX questions. Thanks!

And here is the error message ( you can find it in the notebook but i wanted to add it just for SEO purposes, in case someone is looking for a solution to the same problem )

Fail
Traceback (most recent call last)

<ipython-input-18-8bd369dd3daf> in <module>()
      1 import onnxruntime as rt
      2 
----> 3 session = rt.InferenceSession("output/model.onnx")
      4 input_name = sess.get_inputs()[0].name
      5 print(input_name)

1 frames

/usr/local/lib/python3.7/dist-packages/onnxruntime/capi/onnxruntime_inference_collection.py in _create_inference_session(self, providers, provider_options, disabled_optimizers)
    356         session_options = self._sess_options if self._sess_options else C.get_default_session_options()
    357         if self._model_path:
--> 358             sess = C.InferenceSession(session_options, self._model_path, True, self._read_config_from_model)
    359         else:
    360             sess = C.InferenceSession(session_options, self._model_bytes, False, self._read_config_from_model)

Fail: [ONNXRuntimeError] : 1 : FAIL : Load model from output/model.onnx failed:Fatal error: AliasWithName is not a registered function/op

I Thank you in advance!

[thiagocrepaldi]

I am working on adding support to ONNX export (with and without caffe2) #4120 and #4153

Hopefully it gets accepted soon

[RajUpadhyay]

If you are still stuck on this, then you can use the export_model.py present in the tools/deploy/ folder of detectron2.
It works well since I've used it.

[bouachalazhar]

I am sorry but I do not quite get the question. Have you not been using detectron2?

If not then yes, since this python file is in detectron2, please clone detectron2, and run setup.py

Follow the installation guide.

If I have misunderstood your question, do let me know.

I have installed detectron2 : 0.6 with python -m pip install 'git+https://github.com/facebookresearch/detectron2.git' and I use it but I don't see export_model.py

[RajUpadhyay]

Ah ok I see, then just clone it. Since you have already installed the detectron2 package which you can check via pip freeze | grep detectron2, just clone the git repo and you can use export_model.py.

[bouachalazhar]

Ah ok I see, then just clone it. Since you have already installed the detectron2 package which you can check via pip freeze | grep detectron2, just clone the git repo and you can use export_model.py.

Ok I do it. Is it the best way for export detectron2 model to onnx ?

[thiagocrepaldi]

Ah ok I see, then just clone it. Since you have already installed the detectron2 package which you can check via pip freeze | grep detectron2, just clone the git repo and you can use export_model.py.

Ok I do it. Is it the best way for export detectron2 model to onnx ?

yes. another way is doing whatever the unit tests do at https://github.com/facebookresearch/detectron2/blob/main/tests/test_export_onnx.py

[bouachalazhar]

Ah ok je vois, alors clonez-le. Puisque vous avez déjà installé le paquet detectron2 que vous pouvez vérifier via pip freeze | grep detectron2, il suffit de cloner le dépôt git et vous pouvez utiliser export_model.py.

Ok je le fais. Est-ce le meilleur moyen d’exporter le modèle detectron2 vers onnx ?

oui. Une autre façon est de faire ce que les tests unitaires font à https://github.com/facebookresearch/detectron2/blob/main/tests/test_export_onnx.py

How use it, can you give example ? If I understand there isn't function for export model when we install detectron2 package.

[4daJKong]

Hello, have you solved your problem? I also face the same problem #4881

[bouachalazhar]

@bouachalazhar Hi, so I just generated a onnx file for mask_rcnn_R_50_FPN_3x.yaml. I will share the command and my environment details and I will also share the onnx file with you.

NOTE: I still recommend using the docker image because I myself use docker because different version requirements can be painful.

Environment: Ubuntu 20.04 Cuda: 12.0 (11.8 works to be honest, I upgraded my cuda for whole another project) Cudnn: 8.6

onnx==1.13.1 onnxruntime==1.14.1 torch==1.13.1 torchaudio==0.13.1 torchvision==0.14.1

Command python3 detectron2/tools/deploy/export_model.py --sample-image image.jpg --config-file detectron2/configs/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml --export-method tracing --format onnx --output ./ MODEL.WEIGHTS .torch/iopath_cache/detectron2/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x/137849600/model_final_f10217.pkl

ONNX file https://drive.google.com/file/d/1UzIHjDM4Qkq_KPh84G9AK6K7vuwqsK7L/view?usp=share_link

I hope this helps you in some way.

It works for me, I think when I used git clone it does'nt work but my first problem it's to convert ONNX to OpenVINO IR(.xml file). Thx for all !

[RajUpadhyay]

Glad it is working.

[YingKn]

@bouachalazhar Hi, so I just generated a onnx file for mask_rcnn_R_50_FPN_3x.yaml. I will share the command and my environment details and I will also share the onnx file with you.

NOTE: I still recommend using the docker image because I myself use docker because different version requirements can be painful.

Environment: Ubuntu 20.04 Cuda: 12.0 (11.8 works to be honest, I upgraded my cuda for whole another project) Cudnn: 8.6

onnx==1.13.1 onnxruntime==1.14.1 torch==1.13.1 torchaudio==0.13.1 torchvision==0.14.1

Command python3 detectron2/tools/deploy/export_model.py --sample-image image.jpg --config-file detectron2/configs/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml --export-method tracing --format onnx --output ./ MODEL.WEIGHTS .torch/iopath_cache/detectron2/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x/137849600/model_final_f10217.pkl

ONNX file https://drive.google.com/file/d/1UzIHjDM4Qkq_KPh84G9AK6K7vuwqsK7L/view?usp=share_link

I hope this helps you in some way.

Hi,RajU. I export onnx model by tracing too,but I can't load this model with Opencv function cv2::dnn::readNet.Have you tried to deploy your model in c++? Thank you in advance!

[RajUpadhyay]

@YingKn Unfortunately no, I haven't. I only needed the onnx file for the tensorrt engine.
But as far as I've seen, loading onnx model with cv2::dnn::readNetFromONNX() can be quite tricky. (for eg, dynamic input shapes dont work with opencv and all)
It would help to see the error but please see the opencv github issues, I have seen alot of issues related to dnn::read.

[YingKn]

@YingKn Unfortunately no, I haven't. I only needed the onnx file for the tensorrt engine. But as far as I've seen, loading onnx model with cv2::dnn::readNetFromONNX() can be quite tricky. (for eg, dynamic input shapes dont work with opencv and all) It would help to see the error but please see the opencv github issues, I have seen alot of issues related to dnn::read.

ok RajU,thank you so much.