I read here Dataflow compiler best practice - General - Hailo Community that if using an ONNX model, it’s recommended to simplify it using the onnx-simplifier
Python package. For my project, I used the ONNX model without simplification. My question is: Does the simplified version perform better on Hailo? What are the benefits of using a simplified ONNX model?
Hi @dgarrido,
The onnx-simplifier
is mostly useful to clean an ONNX model from unnecessary operators that may cause incompatibilities with the Parser.
An example of this are those operators present in models exported with dynamic shapes (e.g. Gather, Shape, etc.): the onnx-simplifier
can be used to compute all the shapes in the graph in advance - given a fixed input shape - and to remove those operators from the model.
If you already converted your model to HEF format, you may see no difference at all.