# machine-learning
- o
OccultMC
09/02/2022, 9:16 AMAre higher batch bad? I’ve done a 500 audio dataset at a lower batch size and the ai seem to over fit so I wanna prevent that - h
hecko
09/02/2022, 9:16 AMi honestly don't know - h
hecko
09/02/2022, 9:17 AMpeople here seem to like overfitting - o
OccultMC
09/02/2022, 9:17 AMWdym? - h
hecko
09/02/2022, 9:17 AMe.g. the polish section has a habit of training for thousands of epochs - o
OccultMC
09/02/2022, 9:17 AMYeah i don’t need to train it that long just to loss sub 0.15 or sub 0.20 - o
OccultMC
09/02/2022, 9:20 AMWhat about training? I’ve heard halfing the learning rate when the Val/loss plateaus can help training - h
hecko
09/02/2022, 9:41 AMyeah definitely do that - g
Gamma Prime
09/02/2022, 9:13 PMI really want an answer to this question. IS THIS VOICE POSSIBLE? - u
(Dawn) Will Draw Fictional Women
09/02/2022, 9:19 PMAAAAAAHHHH WHAT THE FUCK - g
Gamma Prime
09/02/2022, 9:28 PMWhat is that? - g
Gamma Prime
09/02/2022, 9:28 PMAre you training a model? - g
Gamma Prime
09/02/2022, 9:30 PMThat's one of my better stonks for Astrotrain. - g
Gamma Prime
09/02/2022, 9:31 PMI don't know whether to keep the data I have or just scrap it all. - h
hecko
09/02/2022, 9:36 PMuhhhh guesses: - your learning rate is hella high (pikachu once set it to an integer :) ) - no base model - the code is just that fucked - g
Gamma Prime
09/02/2022, 9:50 PMMy models' graphs have never looked like that. - g
Gamma Prime
09/02/2022, 9:50 PMThat's way too much orange. - g
Gamma Prime
09/02/2022, 9:51 PMI usually have too much purple in mine when they fail. - u
(Dawn) Will Draw Fictional Women
09/02/2022, 9:51 PMexhibit a is accurate - h
hecko
09/02/2022, 9:51 PM👏 - u
(Dawn) Will Draw Fictional Women
09/02/2022, 9:54 PMi did 95e-3 instead of 95e-5 - g
Gamma Prime
09/02/2022, 9:58 PM95e? - g
Gamma Prime
09/02/2022, 9:59 PMIsn't that stupidly high - g
Gamma Prime
09/02/2022, 9:59 PMOr am I misunderstanding this stuff still? - h
hecko
09/02/2022, 10:01 PM95e-3 = 95 * 10^-3 = 95 * 0.001 = 0.095 - h
hecko
09/02/2022, 10:01 PMor 95 times the default learning rate - h
hecko
09/02/2022, 10:01 PM95e-5 is 5% lower than the default - m
matty1406
09/02/2022, 10:03 PM - g
Gamma Prime
09/02/2022, 10:07 PMSo I should be raising the numbers after the - sign? - h
hecko
09/02/2022, 10:11 PMwell it's really decreasing the number after the e sign