SparseInst export to ONNX input data format

[LukasMahieuArinti]

Hi,
I tried exporting the weights of SparseInst (GIAM) that are in this repository to ONNX format using export.py with the following command (I assume I need to use this command? The documentation reads 'use export_onnx.py', but there's no export_onnx.py in the current branch of this repository).

python export.py --config-file configs/coco/sparseinst/sparse_inst_r50_giam_aug.yaml --opts MODEL.WEIGHTS weights/base_giam.pth INPUT.MIN_SIZE_TEST 512

This leads to the following issue:

  File ".../yolov7/yolov7/modeling/meta_arch/sparseinst.py", line 95, in <listcomp>
    images = [x["image"].to(self.device) for x in batched_inputs]
IndexError: too many indices for tensor of dimension 3

Any ideas on how to fix this?
Thank you.

[lucasjinreal]

It was renamed. An onnx file already generated if I am guess it right. Just ignore the error.

[LukasMahieuArinti]

I actually want to run the onnx conversion on my custom trained model as well, which is why I'm looking to get the error fixed :/
(Sidenote; I can't download the pretrained SparseInst ONNX file from google drive either, it seems to have been removed)

[lucasjinreal]

onnx is int8 which seems not work at this moment.

[LukasMahieuArinti]

Okay, thanks for the reply. I'll try to look for a fix myself then, I'll let you know if I manage to find any.

[LukasMahieuArinti]

It looks like onnx export work fine, it's the check that happens afterwards that fails. This happens because the forward pass through the Sparseinst network expects different inputs in training than evaluation.
To fix this, just comment out this check in export.py or change this line and this line to something like if self.onnx_export:, so that the forward pass goes through those specific preprocessing steps as well during testing.

[lucasjinreal]

the onnx exportation already finished in torch.onnx.export, the rest of code is just test onnx inference logic, this is not necessary and definitely could fail. Just ignore..

[mauricewells]

Then, how can I test onnx inference logic without the rest of code?

[jackiewangCV]

I am asking the same question as mauricewells, as well
Please give me the solution