Demo can be exported to ONNX but other pretrained models cannot about detr HOT 9 OPEN

ONNX export would be really useful. Unless there is a way of converting torchscript to ONNX.

from detr.

Hi,

Thanks for your request.
Adding support for torchscript was on our TODO list, so I've just sent a PR making the torchhub models support torchscript #51

We don't currently have plans to add support for ONNX though, but we would welcome PRs making it work (as long as the code still stays readable and simple to understand).

from detr.

@fmassa
Thanks for the link to your PR!
I'll look into adding ONNX support.

from detr.

The model detr_resnet50 can be converted into onnx, but its output is all nan, I tested NestedTensor in another project, its inference result in onnxruntime is also different compare to pytorch's. After removing NestedTensor, its result is equal to pytorch's. I guess the NestedTensor here cause this problem. Is there some suggestion to fix it?

from detr.

Hi all, I think I have fixes the above problem I mentioned. I retrieved the code in torchvision's faster-rcnn repo, and I found the purpose of this function is aimed at resolving this problem. My fix is here, I will test if it can help to resolve the onnx inference problem in DETR.

from detr.

Hi @fmassa

I take a quick fix of nested_tensor_from_tensor_list() according to the implementation of torchvision's to make it supported by ONNX tracing, Now the inference results in onnxruntime is consistent with pytorch. And here is my modification of _onnx_nested_tensor_from_tensor_list().

Can you check my modification, if it is OK, I would like to submit a PR :)

from detr.

@zhiqwang your implementation looks good, can you send a PR?

from detr.

Hi @fmassa , Of course, It's my pleasure.

from detr.

So it it supported now?

from detr.