Ending epochs for Res50？

mlcommons / training_policies

Issues related to MLPerf™ training policies, including rules and suggested changes

https://mlcommons.org/en/groups/training

Apache License 2.0

92 stars 66 forks source link

Ending epochs for Res50？ #371

Closed frank-wei closed 4 years ago

frank-wei commented 4 years ago

Looks like the doc does not document this or I missed to find it. Look at last 0.6 submission, this value varied in different submission.

bitfort commented 4 years ago

I believe this is captured by: resnet lars lars_opt_learning_rate_decay_steps unconstrained

In the HP table.

AI - Make this clearer in the table. (This will be part of a HP name overhaul).

ppwwyyxx commented 4 years ago

For closed division, I'm wondering what's the reason to allow different submissions to train for different number of epochs? If one submission managed to train for fewer epochs than others by tuning parameters such as learning rate, wouldn't that be against the spirit of "closed division"? I saw that submitters can look at parameters of each other and adapt, but why start this game?

Thanks,