LSA problems
-
Trimming SPSBEAM/B takes too long (~2s)
-
Performance profiling on LSA NEXT
- Kevin will implement passthrough deivice
-
Train with complex LR scheduler to avoid local minimums
- Exponential LR first on 1e-2 → 1e-6, then cyclic LR with exponential annealing of peaks
-
Full SFT to LHC dataset now training on Olympus (still)
-
Pass through filter before sending to LSA