Setting opt_level doesn't optimize

Questions
1

I compiled Resnet50 with opt_level = 0, 1, 2, 3, but the time costs have little difference. So what’s the problem? Can anyone help me?

2

Probably you may need to try profiling each operation on tvm graph runtime.
Ref. https://github.com/dmlc/tvm/blob/5b5465b59e1d5f943a46db84d43076a93d8a7003/src/runtime/graph/graph_runtime.cc#L56

3

Does TVM support any Python API to profile each operations on graph runtime?

4

Which device are you using?

5

A personal PC with GTX 1060. Code is something like this:

        with nnvm.compiler.build_config(opt_level=3):
            graph, lib, params = nnvm.compiler.build(
                sym, target, shape_dict, params=params, dtype=dtype)

where target = 'cuda'

6

Ref. https://github.com/dmlc/tvm/pull/1378/

TVM debugger is on the way to get some profiling information.