Is there any api doccuments about the Dataflow Compiler?

General
1

In fact, I just trying to convert an ONNX file to an HEF file with my CUSTUM AI MODEL, and this model has a completely different structure from the official models provided by Hailo (i.e. not the yolo series).

I could understand if the result indicated that the structure is too large to fit on the Hailo 8L due to hardware specifications.
However, before that procedure, when I tried to compile it using the model zoo API, I received a message asking to provide the model structure information in YAML format.

I searched for related documents but couldn’t find any appropriate examples or documentation. I think it might be possible to directly optimize using the Dataflow Compiler.

Any comments or recommended documents related to this would be appreciated.

Thank you for reading.

JSKang

2

Hey @happistday,

You can find the documentation on DFC usage here: DFC Documentation.

Additionally, there’s a quick guide available in the community: Creating Custom HEF Using DFC Model Zoo.

If you have any further questions, please don’t hesitate to ask.

3

Thank you for your sharing document!!

the second one is that just what i found.

Unfortunately, when i follow the procedure, i found the new error types as bellows:

onnx.onnx_cpp2py_export.checker.ValidationError: No Op registered for SimplifiedLayerNormalization with domain_version of 14

==> Context: Bad node spec for node. Name: /model/layers.0/input_layernorm/LayerNorm OpType: SimplifiedLayerNormalization

But, i totally thanks for your sharing, and i will tried to fixed it out.
(p.s.: is there any version to fix this functionality?
currently, i just load the onnx model and tried to changed the problem layers to normal one… (i.e. simplified layer norm to layr norm…)

have a good day. Thank you :smiley:

4

Hi @happistday,
From the error it seems that the model contains a SimplifiedLayerNormalization. But, it seems that this type of layer is not supported by ONNX itself (I also saw a thread about it here: LayerNormalization operator is not in onnx domain · Issue #7573 · microsoft/onnxruntime · GitHub).
The error message comes from the onnx checker, which make sure that the ONNX is valid.
So, this is an ONNX issue and not a Hailo SW issue. From what I saw online this is still an open issue.
You can try using LayerNormalization instead, which is supported with ONNX opset version 17.

Regards,

1 Like
5

Thanks for your comments.
currently, based on your recommendation,
i tried t optimized the onnx model with opset 17,
(unfortunately, the memory consideration issues are happen)
i new that this issues will be considerd in onnx community.

thank you for your comments, and i will be back with the hailo related issues.

have a good day :smiley: