Clarify handling of Non-converged runs

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

Closed bitfort closed 4 years ago

bitfort commented 4 years ago

Clarifying the rules. We believe that runs which don't converge should be counted as infinite runtime when computing the Olympic Score for MLPerf.

bitfort commented 4 years ago

SWG:

We clarify that this is everyone's understanding and that this should be codified in the rules.