Work-in-progress notes about cascade
I started using this dataset with ADAMW, but then started using Adafusion/adafusion, which was great. Then I tried that out on cascade.
Turns out, LR of 6e06 did very little. So I kept increasing it until I saw noticable differences between epochs. Then I found the reasonable upper bounds.
THEN I decided to damp it down a bit, with EMA use. Has to be CPU based, with default attention and no gradient saving, or it wont fit in memory.
That seems to be the best combination for cascade training I've found so far. Will upload a specific OneTrainer JSON when I feel like I've found "best" results