There is a problem with autotvm, when I convert the weight and use debug_skip_region in the schedule

#1

I use the conv2d_alter_op pass for weight transform. I used debug_skip_region in the schedule. On a version around April 20, autotvm can work normally, but after I update to the latest version, autotvm cannot tune the log, and many errors will occur during the tune process.

#2

Use debug_skip_region will cause inaccurate autotvm measurement, recently some fixes were merged: https://github.com/apache/incubator-tvm/pull/5684

Is this related? @cchung100m

#3

Hi @kevinthesun

Thanks for the notification. I’m not sure it is related to our PR or not.

Hi @123

Can you provide the error messages and snippet code for demonstration?

#4

thanks for your reply! I am not sure whether it is the effect of debug_skip_region,I reset the n_trial, early_stopping, although there will still be errors, but it can run out of performance.