Nvvm or relay, which to use when autotvm under gpu

  1. when i use: net, params = nnvm.frontend.from_onnx(onnx_model), it shows tvm.error.OpNotImplemented: The following operators are not supported for frontend ONNX: InstanceNormalization

  2. when i use: net, params = relay.frontend.from_onnx(onnx_model, shape_dict), it can successfuly run get_network(), but abort in autotvm.task.extract_from_graph()

Use Relay with autotvm.task.extract_from_program()