Issue running examples with custom model
Trying to run bytetrack_plus_demo generic inference demo with custom model. Model was converted from onnx to nef successfully, but using NEF_Parser_lite produces an error.
AssertionError: Invalid setup.bin format
Ignoring/commenting out the error and trying to proceed produces another error later on.
Error: inference failed, error = Error raised in function: generic_raw_inference_bypass_pre_proc_receive. Error code: 30. Description: 30
This is the onnx model used at the beginning. Unsure if this was a conversion issue or a KL720 issue.
The discussion has been closed due to inactivity. To continue with the topic, please feel free to post a new discussion.
Comments
@Clarence
Hi Clarence.
I'm not sure which version of Kneron Toolchain you are using, would you mind letting us know.
This should be a problem with the model conversion, in KL720, the main reason for this model error is because the value info of the model is dynamic input instead of fixed input.
So you can try to change the model unk__32 to 1 and it should work.
Thank you for the reply. I have changed the input batch size to be fixed at 1, as shown below.
However I am still receiving the invalid setup.bin error.
Again if I ignore and comment out the assert line...
And then again I receive the following...
@Clarence
Hi Clarence,
The reason for the error may be that your onnx model is missing "value_info" between layers.
You can try to run onnx's "shape_infernece" or "polish_model", these should help to calculate the middle shape.
I have updated the model to look as above, but I get the same errors as before.
@Clarence
Hi Clarence,
I may need to confirm with you that the model you are using (onnx, nef) and the Scrtips, and was wondering if you could provide them.
I'm guessing you're probably using the byte-tracing code you provided earlier in the contest since the byte-tracing code in PLUS shouldn't have "NEF_Parser_lite.py" in it now.
For the current Kneron PLUS you may refer to: https://www.kneron.com/tw/support/developers/