CRS08 Unsupervised tuning notes
Created on June 28|Last edited on June 28
Comment
Some basic runs
Run set
13
- Coarser unsupervised tuning works, as before
- However, it's too slow (30 minutes???).
- Larger batch size (4 -> 32) helps the human-pt model, TBD for subject-pt.
- Accelerated LR does help with time but kills peak performance, we'd rather just wait?
Still waiting to test actually broad pretraining (with Will's, still motor, but non NDT tests).
Add a comment