Hello,
During autotuning on android device, very often device is restarting, so the tuning process is very uncomforable. I have question, how to check which config causes broken run? How can I filter out broken configs? Or how to configure autotuning to avoid problematic configs?
Do you use AutoTVM or AutoScheduler? Are you tuning on CPU or GPU?
I also got into that problem and theoretically we can restrict the search space in case of AutoTVM but not sure what we can do with Ansor. On the other hand it must be less important for Ansor since the nature of schedule mutation will be based on previous working schedule and we must not go to the danger area where execution will be too long.
In the same time I do not see that limiting of search space and filtering out for AutoTVM will be generic and can be reused on any hardware and schedules. It must be a process of collecting of statistics and limiting of knobs based on collected data
Hi,
I’m not sure, but I think I have been using AutoTVM (for now I’m following these tutorials: Auto-tuning a Convolutional Network for Mobile GPU — tvm 0.9.dev182+ge718f5a8a documentation and Deploy the Pretrained Model on Android — tvm 0.9.dev182+ge718f5a8a documentation). Does AutoScheduler will speed up my tuning? I’m using GPU via OpenCL.
I think I have solved my problem in other way. I was using rooted phone with custom build so the problam was phone’s software. Now I’m using market-phone, and problem disapear.