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
92 stars 66 forks source link

SSD Reference for 1.0 #385

Open bitfort opened 4 years ago

bitfort commented 4 years ago

Tracking for the SSD reference for 1.0

bitfort commented 4 years ago

SWG:

Investigated why some submissions reach a higher average quality score at the 95 epoch. It appears some submissions were using a non-reference backbone and a bias=0 initialization.

  1. We should establish a backbone conversion script for SSD (pytorch to tensorflow).
  2. Establish the expected behavior of the SSD accuracy and have compliance checkers look for potential issues.
  3. Rules should state checkpoints and backbones are provided in top 2 frameworks (currently pytorch and tensorflow) and an automatic conversion script is provided between them.
  4. You must use one of these two backbones unless you show the SWG why this isn't possible and propose making another backbone available.