Re: [dmlc/tvm] [RFC][AutoTVM] Selective Tuning (#4188)

[Cody Hao Yu <no...@github.com>]

While I am testing if tuning more trials could make the result more intuitive, I would like to first ask for the feedbacks about the naming. Here are my thoughts based on Tianqi's comments.

- select.py -> pass.py
As suggested, this module is more like a pass over a set of tasks, so we can treat it as a pass. I can implement it as a pass table so that we can add other passes in the future.

- Select representative tasks. Origianl: `autotvm.task.mark_depend(tasks)`
With the pass implemented above, this API becomes `autotvm.task.pass.FindReference(tasks)`, meaning that this pass is going to find the reference task for each task.

- task.depend -> task.cfg_ref_task
`cfg_ref_task` points to the task that we can refer its tuned configs when tuning. 

- tuner.depend_mode -> tuner.cfg_ref_mode
`cfg_ref_mode` is a string in the `<mode>-<cfg>` format. `<mode>` can be either "only" or "start"; while `<cfg>` can be either "topN" or "N%". Here are some examples:
    - "only-top10": Stop tuning after trying the top 10 configs in the `cfg_ref_task`.
    - "start-5%": First try the top 5% configs in the `cfg_ref_task` and back to the normal tuning.
    - The default is set to "only-top10", meaning that we will stop tuning after 10 trials if the task we are tuning has a tuned reference task.

@kevinthesun @eqy @icemelon9 please let me know what you guys think about these names, and you're welcome to propose better ones. Thanks.


-- 
You are receiving this because you are subscribed to this thread.
Reply to this email directly or view it on GitHub:
https://github.com/dmlc/tvm/issues/4188#issuecomment-548992976