What batch size is appropriate for a 1 minute and 30 second dataset in ov2? | AI HUB | Page 1

mighty crow Jan 15, 2024, 7:02 PM

#

8? 4?

queen basalt Jan 15, 2024, 7:02 PM

#

Go for 4

#

So you get more steps per epoch

#

And therefore more training time

#

But better model overall

mighty crow Jan 15, 2024, 7:03 PM

#

thx

#

Solved

buoyant summit Feb 9, 2024, 8:59 AM

#

Epochs?

queen basalt Feb 9, 2024, 12:36 PM

#

Just go for 500, every dataset length

buoyant summit Feb 9, 2024, 12:41 PM

#

ok

pine leaf Feb 9, 2024, 2:00 PM

#

200-300 epochs for ov2

queen basalt Feb 9, 2024, 2:01 PM

#

I recommend 500 so it's not undertrained

It will most likely overtrain at 300-400 epochs anyway, especially on batch 4

pine leaf Feb 9, 2024, 2:10 PM

#

500 for default pretrain tho. I got overtrain sign at 200-300 epochs, or got optimal epoch at that range for ov2 with batch size 8.

queen basalt Feb 9, 2024, 2:13 PM

#

either way, you kinda have the risk of undertraining at 500

its just a placeholder target that can be changed, its not what you should necessarily use, thats why i say 500

#What batch size is appropriate for a 1 minute and 30 second dataset in ov2?