The math with training images, steps, and epochs doesn't sum up right to me, can you clarify?

#34
by alexds9 - opened

You are stating: "It was trained over Stable Diffusion 1.5 with +60000 images, 4500 steps and 3 epochs.".
If you used 60K training images, and trained for 3 epochs, you need 60K * 3 = 180K steps of training steps.
With 4.5k steps and 3 epochs, the amount of training images should be 4.5k / 3 = 1.5k images.
Obviously, something is not summing up in the stated numbers.
Can you please clarify what exactly am I getting wrong, is there an error in the original statement?
Also, it's very interesting to know how you were able to scrape 60K images from Midjorney, and how long the training took.
Thanks!

alexds9 changed discussion title from The math with training images, steps and epochs doesn't sum up right to me, can you clerefy? to The math with training images, steps, and epochs doesn't sum up right to me, can you clarify?

doesn't it mean the batch size is 60000 * 3 / 4500 = 40?

@haowang1013 Maybe, is it possible to do batch size 40?
Hopefully, @prompthero can clarify it and other settings of the training.

@haowang1013 Maybe, is it possible to do batch size 40?
Hopefully, @prompthero can clarify it and other settings of the training.

The code certainly allows it, so it's just a matter of making sure it fits in the GPU memory and there're many tricks to make it happen.

But yeah having the original creator clarify the settings is def the best.

Sign up or log in to comment