How convert from OpNode to FunctionNode in Relay IR?

#1

Hi,

At frontend, how to convert OpNode to FunctionNode? Facing the below error when trying to set function attributes with “func.with_attr” at frontend

  File "/data1/workspace//incubator-tvm/src/relay/backend/graph_runtime_codegen.cc", line 407
TVMError: Operators should be transformed away; try applying the fuse_ops transformation to the expression.

Could see the testcases are using relay.Call(), can anyone please help me on how to do this at the frontend file (eg: tflite.py) ?

#2

You may use eta expansion to convert a primitive op to a function.

#3

Could you please give me a sample code?

#4

In most cases, we cannot simply convert opnode to function nodes, as they are primitive ops, you can however, wrap the op with function that calls into this op.

#5

@tqchen, i will explain my specific problem. I am trying to use the external c codegen from tflite frontend. I tried to check with below method as shown in some testcases shown on external codegen.

    func = _function.Function(analysis.free_vars(outputs), outputs)             
                                                                                
    func = func.with_attr("Primitive", tvm.tir.IntImm("int32", 1))              
    func = func.with_attr("Compiler", "my_compiler")                            
    func = func.with_attr("global_symbol", "my_compiler_0")                     
                                                                                
    #mod = IRModule.from_expr(call)                                             
    mod = IRModule.from_expr(func)

But the relay.Call() call is not able to perform at this stage and am facing the issue with OpNode. Can you please suggest a method to resolve this issue?

#6

cc @zhiics might have a suggestion, what is the content of the function? a function must contain a sub-graph/expression, if you want to use an op you must call it with inputs.