[BUG] Can't convert segformer ONNX Model to Blob Files

[yuchenlichuck]

Check if issue already exists

• Google it (e.g. error xy github luxonis depthai)
• Check troubleshooting in documentation.

Describe the bug
A clear and concise description of what the bug is.

Minimal Reproducible Example
Append the MRE to the bug report, instructions here

Expected behavior
I tried to convert segformer onnx model to blob files in http://blobconverter.luxonis.com/.
But I failed many times and there is no error output. Could you help me know what is wrong with it? Thank you.

Screenshots

Pipeline Graph

Please also provide a screenshot of your pipeline using the DepthAI Pipeline Graph.

Attach system log

• Provide output of log_system_information.py

Additional context
Add any other context about the problem here.

[brmarkus]

Can you provide a pointer to where to get exactly the same model you tried to convert/compile, please? Is this an "original" model, or your model, have you changed the model?
Have you tried converting/compiling other models - just for consistency-check if the environment is installed properly, the model fitting (like supported OPS-level, etc.)?

[Erol444]

Model (from forum thread - deleted):
https://drive.google.com/file/d/1lo8wW020YS5nTpZ8iAd3HWRUw2uV67CK/view?usp=sharing

[Erol444]

@yuchenlichuck as Markus requested, please also share model repo, and how/where you got the onnx.

[yuchenlichuck]

I got the model from Hugging face segments-tobias/segformer-b0-finetuned-segments-sidewalk

from transformers import AutoFeatureExtractor, SegformerForSemanticSegmentation
model = SegformerForSemanticSegmentation.from_pretrained("segments-tobias/segformer-b0-finetuned-segments-sidewalk")

then I use pytorch onnx to generate the onnx files.

import torch
dummy_input = torch.rand(1, 3, IMAGE_HEIGHT, IMAGE_WIDTH)


torch.onnx.export(
    model,
    dummy_input,
    onnx_path,
    opset_version=11,
    do_constant_folding=False,
)

[Erol444]

@tersekmatija could someone from ML team quickly evaluate whether this would be possible?
Thanks!

[tersekmatija]

I've tried quickly exporting this locally with the latest API and it seems that the compiler gets stuck. No error or anything, it just times out. I would suggest opening and issue on OpenVINO GitHub and linking it here. From what I've observed, transformers are not really optimized for MyriadX.

[yuchenlichuck]

Okay, thank you for sharing this information with me!