How to tune GNN models in TVM?

Questions
#1

Hello everyone,

I am trying to tune a GNN model (feastconv_opset16.onnx) using AutoTVM, but I get this error during task extraction:

RuntimeError: flop estimator fails for this operator. Error msg: the length of axis is not constant. Please use cfg.add_flop to manually set flop for this operator.

Here is what I tried:

onnx_model = onnx.load(“feastconv_Opset16.onnx”)

input_shape = { “x”: (2708, 1433),
“edge_index”: (2, 2708)
}

mod, params = relay.frontend.from_onnx(onnx_model, shape=input_shape)

target = “llvm”

tasks = autotvm.task.extract_from_program(mod, target=target, params=params)

How should I handle this error? Do I need to manually set cfg.add_flop, and if so, how?Or is there a better solution to fix this?Can GNN models be tuned efficiently with AutoTVM? Or is there a better approach for optimizing GNN models in TVM?

Thanks!