Skip to main content
awfidius
Projects
pure-transformer
Reports
Loss depends on batch?
Log in
Sign up
Share
Comment
Star
Share
Comment
Star
Loss depends on batch?
Batches are randomized, so loss should not depend so strongly on batch ID... debugging.
Andrew Fitzgibbon
Created on May 27
|
Last edited on May 27
Comment
Section 1
batch/1000+.2, loss
batch/1000+.2, loss
0
5k
10k
15k
20k
25k
30k
Step
1
2
3
4
5
1-bit-grad
1-bit-grad
Run: 1-bit-grad
1
Add a comment