Training Steps questions

#21
by Linaqruf - opened

Hey, great model and thanks for sharing the hyperparams! I'm scaling the learning rate to use with lower batch size and it help my SDXL training. I'm considering to fine-tune with this model soon, but I have some question.

When you mention "Steps: 251000", are those pure steps without considering the batch size and grad accumulation? Or have they already been divided by the batch size/grad steps, like around 1960 steps? Also, it'd probably be helpful if you could also provide the number of epochs.

Thanks!

Segmind org

Hello! Those are pure steps. We overall did roughly 1.5 epochs on the data. Do you have any specific finetunes in mind?

Oh, okay. Thank you for the information. I plan to fine-tune an anime model based on SSD-1B. I am currently gathering 1 million images from Danbooru and curating generated images from Nijijourney. However, I might try training with a smaller dataset first.

Segmind org

Good luck on the finetune!

Sign up or log in to comment