PROPOSAL Add 3D Unet for v1.0

bitfort commented 4 years ago

There is a proposal to add 3D Unet for 1.0; tracking discussion here.

johntran-nv commented 4 years ago

For 10/8/2020, Pablo will attend and give an update, along with a call for submitters to form a working group.

bitfort commented 4 years ago

SWG:

Forming a Task Group for Next Week (Oct 15).

bitfort commented 4 years ago

SWG:

Update from Pablo;

Currently 3 participants in the task group for the benchmark, meeting for 3 weeks (thursday meeting)
Shared differences between training and inference version of the model
Instance norm is a problematic layer for normalization when scaling for training (do we want batch norm?)
Leaky relu v. relu doesn't make a difference, making to relu may be simpler for everyone
Changing instance norm shows a difference, distributed batch normal preforms better at scale
Making these changes is a trade-off; batch norm preforms more poorly in single-GPU use case.

johntran-nv commented 3 years ago

This is already well under way for v1.0. Closing, since I don't see a benefit from keeping this issue open.

mlcommons / training_policies