Hello!
I know that “Relay Frontend” make graph optimization by using relay.transform.pass like FuseOps or ConstantFolding.
But I want to block all optimization pass for my experiment. So I adjusted “opt_level” when building “Relay” module like below.
with relay.transform.build_config( opt_level= 0 ):
graph_json , Lib , Params = tvm.relay.build_module.build( mod , target='cuda' )
The mod have layers like below.
I want to build network without graph level optimization like the order above in my picture, but when i build module they automatically fuse some layer even when opt_level is zero.
So my question is if there are any function or options that can block all graph_level_optimization?