Centernet: Provide ONNX models where possible

[gyenesvi]

Can you please provide ONNX models of those network that are possible to export to ONNX? It seems to me that at least some variants of CenterNet for object detection should not contain any special operations that could not be exported to ONNX, so it seems doable. However, it seems that multiple people (including myself) have tried to load the models available here and export them to ONNX and failed.

My attempts failed at loading models due to unwanted dependencies in the code base. I tried to import a single detector (CtdetDetector) but failed because importing that model depends on all other models in the code, which require cuda (I don't even have cuda GPU) and external nms op to be compiled, which is not required for this detector network as far as I understand.

[xingyizhou]

I honestly don't know what ONNX is and don't have a good motivation to learn this ...

[gyenesvi]

ONNX is a format for exporting your models so that others can use it outside of PyTorch as well. I believe that should be a good enough motivation, you probably want your research to be accessible to others as well. In the present format, your models cannot even easily be loaded in PyTorch as I explained above.

[bhack]

It is also partially related to JIT/Torchscript your models.

[Markusgami]

Just change the ouput of the network from dict to list, then it will work(without deformable conv)

[jinfagang]

@Markusgami How do u remove these deformable conv?

[sezan92]

don't have a good motivation to learn this ..

making an open neural network model which can be used across all frameworks is enough of a motivation

[jinfagang]

I found there were people already done this, by exporting centernet to onnx

[sezan92]

I found there were people already done this, by exporting centernet to onnx

can you give the link?

[jinfagang]

http://manaai.cn/aicodes_detail3.html?id=49 They have also accelerated centernet using TensorRT

[PankajJ08]

http://manaai.cn/aicodes_detail3.html?id=49 They have also accelerated centernet using TensorRT

How to access mana ai code outside china?

[ckcraig01]

There will be always engineering people in the industrial society do the onnx converting, I agree to have the author focus on the research that he is good at (and that's reason why pytorch exists)

[EnricoBeltramo]

Here a example of conversion of CenterNet in ONNX:
https://github.com/CaoWGG/TensorRT-CenterNet