How to get graph representation of Relay program?

johnmattew490 · February 10, 2020, 8:00pm

Hi All,

I am wondering how can I get the graph representation of relay program ? I’d like to get the graph representation in some kind of data structure where I can modify the graph in a way that I want.

cv_ruddha · February 10, 2020, 8:10pm

You get the graph as a return value from the relay.build() function along with library and params.

johnmattew490 · February 10, 2020, 9:00pm

Thank you @cv_ruddha, and may be I should pose my question more clearly.

I’d like to modify the Relay IR, and I believe Relay IR is represented as a Relay Module.

I used relay.analysis.post_order_visit() function to traverse the Relay IR, but I can not modify the Relay IR nodes and create a new Relay IR module. So, my questions is how can I traverse convert Relay IR to another Relay IR?

cv_ruddha · February 10, 2020, 9:08pm

The graph you get from the relay.build is in a json format, you can make changes to the json string and then deploy it onto any hardware target. Once you make a change in the json string you are essentially changing the Relay IR.

johnmattew490 · February 10, 2020, 10:20pm

Thank you @cv_ruddha. truly appreciate the prompt response.

Q1) It looks like we can do all Relay Level IR optimizations on the JSON file. Then, why we need all the RELAY IR? Isn’t the Relay IR is supposed to be the medium to do optimizations? Everything I read so far says that we need RELAY IR because it allows us to do optimization. Please correct me and help me understand.

Q2) I think i’d still want to do get some kind of representation of Relay IR into some kind of graph so that I can do some optimizations. I am not sure if JSON file is the right place to do that optimizartion.

cv_ruddha · February 10, 2020, 10:27pm

You will have to trace the functions being used to optimize relay IR. For example, common sub expression elimination is an inbuilt relay transform. See where and how it is implemented. I know that to create your own transform you need to register your function using the TVM_GLOBAL_REGISTER Macro. I haven’t worked much on building my own transforms yet so I am not sure how that works. If anyone else in the forum knows about it please attach it to this thread.

johnmattew490 · February 10, 2020, 10:55pm

Thank you @cv_ruddha. Yes, now I think we are on the same page.

I have looked at the examples and such as common sub expr elimination from src\relay\pass. I believe I can write my own pass following common sub expr elimination and register my pass, but I am looking into achieving this inside PYTHON.

Thus, how can I get a graph representation of RELAY program inside a python ?

Thank you again for the prompt response and sharing your opinion.

MarisaKirisame · February 11, 2020, 1:18am

github.com

apache/incubator-tvm/blob/master/python/tvm/relay/transform.py#L1041


            return _wrap_class_function_pass(pass_arg, info)
        if not isinstance(pass_arg, (types.FunctionType, types.LambdaType)):
            raise TypeError("pass_func must be a callable for Module pass")
        return _transform.MakeFunctionPass(pass_arg, info)


    if pass_func:
        return create_function_pass(pass_func)
    return create_function_pass




@function_pass(opt_level=1)
class ChangeBatch:
    """
    Change the batch size.


    Parameters
    ----------
    data: Dict[relay.Var, int]
      A dictionary of all the params to change.
      The keys are all params, and the values are which dimension hold the batch.

johnmattew490 · February 12, 2020, 2:21am

Thank you @MarisaKirisame

I saw this python file, but I am looking for ways to get graph representation of Relay IR? How can I get a graph representation of Relay IR and make some modification on that Relay IR ?

Or is it not possible to get the graph representations of Relay IR?

MarisaKirisame · February 13, 2020, 12:34am

the relay IR is a graph representation. you can print(x) to see the text format. call dir(x) to check it’s member whih can be access through dot. If you want the API for all IR representation, see the expr.py file.

johnmattew490 · February 13, 2020, 1:04am

Thank you @MarisaKirisame. Super appreciate the response.

What I am trying to do is to get a REAL graph where Vertices are nodes, and also with edges which tells me dependenccies. Does Relay IR is DAG?

For example, let us analyze following relay program.

 # Create a simple Relay program.
    c = relay.const(2.0, "float32")
    x = relay.var("x", "float32")
    y = relay.add(c, c)

    func = relay.Function([x], y)
    mod = relay.Module({"main": func})

Here is what mod, print(mod), dir(mod) gives if in the python console. And as I can see, it does not give any DAG info of Relay IR?

>>>mod
IRModuleNode( {GlobalVar(main): FunctionNode([Var(x, ty=TensorType([], float32))], (nullptr), CallNode(Op(add), [Constant(2.0), Constant(2.0)], (nullptr), []), [], (nullptr))})

>>>print(mod)
v0.0.4
def @main(%x: float32) {
  add(2f, 2f)
}

>>>dir(mod)
['functions', 'global_type_var_map_', 'global_var_map_', 'type_definitions']

MarisaKirisame · February 13, 2020, 1:53am

func is your RelayIR. It is a graph. you can recover func from Module via mod.functions.

johnmattew490 · February 13, 2020, 2:40am

Thank you. I posed an example here showing what I am trying to achieve: How to do Relay to Relay transformation by traversing each node?

Using the “func” or module, I am still can not write a simple relay to relay program transformation after visiting each node.

AnandGokhale · May 25, 2020, 11:39am

Hi, I was trying to use this for a different application, but I seem to be losing information about the ops while doing so, for example, while performing conv2d, or maxpool2d, I lose the information about padding, stride. Is there a graph representation exposed that doesnt ignore this information?