[Auto-TVM] The train result is not better than pretained result

Questions
1

I have used auto tvm to tune for my device. However the result is not better than pretrained configuration in the tophub. My device is not ra3b, but when I use ra3b’s pretrained configuration, I can run mobilenet in about 350ms，but I use auto tvm to train for my device, I only can run mobilenet 450ms. It is not normal. My train script is the same as docs of tvm. @merrymercy could you give some advices? For example n_trials from 1000 to 1500, early_stopping from 200 to 400? time_out = 10 becomes time_out=60 ?These can help performance result? Thanks in advance.

2

If you have enough time budget, you can set
n_trials to 1200 and early_stopping to 600. Most tasks finish due to early stopping, so early stopping is a more important parameter.

Timeout is used to filter out too slow configurations. For your current case, it is okay. You needn’t change it.

3

Thanks for replying. If early stopping is important, we set it equal to n_trial is better? If we have time budget.

4

Yes, you can set it to None.

5

Thanks. I will try it.