Generation of .blob file using modelconverter and blobconverter

AyushmanChoudhuri

Hello!

I am currently using .blob files for the yolov6r3 model and have been using the online luxonis tools interface to convert models from pytorch to .blob format. I want to migrate to using the luxonis python packages to take care of the conversion process. I have tried the following:

Used the legacy blobconverter cli tool to convert a yolov6r3 onnx file to a .blob file. However, this does not generate the config file that is auto generated using the online interface.
The blobconverter also only takes in the yolov6r3 onnx file using the cli and not the pytorch weights. Is there a way to convert the pytorch file directly to the .blob file using the cli.
When using the latest modelconverter package, only the superblob file in the NNArchive format is generated even when passing superblob=false in the conversion script. Is there anyway to use the latest modelconverter package to generate .blob files, or is the .superblob the only format supported ?

KlemenSkrlj

Hi,
I belive we are addressing these issues now in this thread but for visibility I'll quickly answer it here as well:
1,2/ Correct, legacy blobconverter is meant for converting non YOLO models. The intended legacy conversion path for YOLO models is to use tools web app which takes in a .pt file and converts it to a blob file.
3/ This is a bug which we are addressing right now (PR here) and will be available in the next release of modelconverter package. Additionally now there is also an option to convert to .blob (instead of .superblob) through Hub UI where you can pick "Convert to blob" in advanced RVC2 conversion parameters.

Best,
Klemen