Replies: 1 comment
-
I changed the steps from 500 to 1500 because I lowered the learning rate from 1e-5 to 2e-6 in the default settings to make sure users won't face excessive overfitting. the value 1e-6 seems okay for the text encoder. Upping the values will accelerate the training but could break the model, so one would have to experiment with different settings. |
Beta Was this translation helpful? Give feedback.
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
Some days (?) ago, some preset values in the training section got updated with new preset values:
UNet_Training_Steps from 500 to 1500 (per 10 images)
UNet_Learning_Rate from (?) to 2e-6 (lowest)
These changes are quite significant, and I would be interested: why and to what effect? And is there a dependency to the "no more than 20 pictures for a portrait" mantra?
And then:
Text_Encoder_Learning_Rate = 1e-6
But the help text reads: keep it low to avoid overfitting (1e-6 is higher than 4e-7)
So - why not changing the preset to a lower setting, what are the pros/cons?
Thanks in advance.
Beta Was this translation helpful? Give feedback.
All reactions