Cannot convert .onnx file to .blob file using python script and luxonis blob converter takes forever

[KennethEladistu]

Hello, currently I am using yolo v8 OBB for this one and want to integrate my model on our Oak-D-Pro, currently I already converted it into ONNX file and is looking for way to convert that .ONNX file into .BLOB file. I already tried 2 ways first is the blob converter I run it for 7 hours and still has not yet converted my file into .BLOB.

I also tried the code which is

import blobconverter

blob_path = blobconverter.from_onnx(
model="C:/Users/KennethAaronEladistu/Documents/PT weights/best(1).onnx",
data_type="FP16",
shaves=6,
)

and this is the output of it

[HonzaCuhel]

Hi @KennethEladistu,

Thank you for sharing this issue. Could you please share with use the source .pt and .onnx files so that we might take a closer look?

Best,
Jan

[KennethEladistu]

Hi @KennethEladistu,

Thank you for sharing this issue. Could you please share with use the source .pt and .onnx files so that we might take a closer look?

Best, Jan

Hello this is the link to my source.pt and .onnx file

https://drive.google.com/drive/folders/1x-IRw6zsS-eKANdP3b352yYiYbQXoY1I?usp=sharing

[HonzaCuhel]

Hi,

Thank you for sharing. I'll look into it and get back to you as soon as I find something.

Best,
Jan

[KennethEladistu]

Hi,

Thank you for sharing. I'll look into it and get back to you as soon as I find something.

Best, Jan

Thanks!

[HonzaCuhel]

Hi @KennethEladistu,

I am very sorry for such a long delay in my response. The model wasn't converted because of an unsupported Sin operation at the end of the model (as highlighted in the screenshot).

That's why this part (post-processing) must be pruned from the model and applied to the output of the pruned model. We are currently implementing the support for the YOLOv8 OBB model in our DepthAI; however, as far as I know, we have a flexible ETA. If you need to use the model as soon as possible, I can export the pruned model for you and give you some pointers to what needs to be implemented to use the model on an OAK device. Another option would be to wait for our official support of this model type.

Again, I am genuinely sorry for the delay in my response.

Kind regards,
Jan