search

mlcommons / training_policies

Issues related to MLPerf™ training policies, including rules and suggested changes
https://mlcommons.org/en/groups/training
Apache License 2.0
93 stars 66 forks source link

Number of runs for Each Benchmark #199

Closed bitfort closed 4 years ago

bitfort commented 5 years ago

How many times should each benchmark be run?

bitfort commented 5 years ago

SWG Notes:

Resnet - 5 (both SGD and LARS) SSD - 5 Maskrcnn - 5 Transformer - 10 GNMT - 10 Minigo - 10

cliffwoolley commented 5 years ago

Note that the Reinforcement learning (Minigo) number should be updated on https://github.com/mlperf/policies/blob/master/training_rules.adoc#11-benchmark-results , where it still says 20.

petermattson commented 4 years ago

Effectively a duplicate of #155.

petermattson commented 4 years ago

On second thought, reopen and make active to serve as reminder to get DLRM and BERT number of runs into rules.

petermattson commented 4 years ago

Duplicate of #297