Create METHODOLOGY-cascade.md
Browse files- METHODOLOGY-cascade.md +13 -0
METHODOLOGY-cascade.md
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Work-in-progress notes about cascade
|
| 2 |
+
|
| 3 |
+
I started using this dataset with ADAMW, but then started using Adafusion/adafusion, which was great.
|
| 4 |
+
Then I tried that out on cascade.
|
| 5 |
+
|
| 6 |
+
Turns out, LR of 6e06 did very little. So I kept increasing it until I saw noticable differences between epochs.
|
| 7 |
+
Then I found the reasonable upper bounds.
|
| 8 |
+
|
| 9 |
+
THEN I decided to damp it down a bit, with EMA use.
|
| 10 |
+
Has to be CPU based, with default attention and no gradient saving, or it wont fit in memory.
|
| 11 |
+
|
| 12 |
+
That seems to be the best combination for cascade training I've found so far.
|
| 13 |
+
Will upload a specific OneTrainer JSON when I feel like I've found "best" results
|