We currently use onnxsim to simplify the ONNX graph before transforming it into an SDFG. The main benefit we currently get from this is that onnxsim applies constant folding, which allows us to handle much simpler cases for the supported ops, such as only having constant inputs for some transformation parameters, like the axes of a reduction.
onnxsim is not a reliable dependency and should be replaced with a constant folding pass that simplifies the SDFG at the same level.
We currently use
onnxsimto simplify the ONNX graph before transforming it into an SDFG. The main benefit we currently get from this is thatonnxsimapplies constant folding, which allows us to handle much simpler cases for the supported ops, such as only having constant inputs for some transformation parameters, like the axes of a reduction.onnxsimis not a reliable dependency and should be replaced with a constant folding pass that simplifies the SDFG at the same level.