Graph tuning a TE computation? te.op to relay.Expr?

Questions
#1

Is it possible to graph tune a te based computation?

Graph tuners expect a relay.Expr, but so far I could not find a way to convert te based expressions to relay expression.

Whenever I use te-based constructs (tensors and ops) I found them to be incompatible with relay constructs, is it possible to convert a te.compute(…) (or any other re.op) to a relay.Expr?

Thanks!

#2

Hi, graph tuning is a graph(relay) level optimization. Usually you need to create a relay op which wraps your te computation to use graph tuning.

1 Like
#3

Thanks for the answer Kevin. What’s the best way of doing this?

#4

I could not find a way to use te.Tensors with relay ops or expressions. What would then be the intended way to wrap te constructs in a relay op?

#5

https://tvm.apache.org/docs/dev/relay_add_op.html