google/edward2

Improve deterministic efficientnet baseline

dustinvtran opened this issue · 2 comments

dustinvtran commented

needs:

  • EMA
  • AutoAugment

More investigation is required to close the gap between the cloud_tpu codebase's results which uses TF estimator and our codebase's results.

dustinvtran commented

There's a recent effort (and success!) by @Hyperparticle to reproduce efficientnet in Keras+TF2. Let's wait until that codebase stabilizes and then assimilate it.

Assuming timelines match, let's also do this before tuning BatchEnsemble (#277).

  • 👍1