Fine tune naver-clova-ix/donut-base with your dataset

#1
by FranciscoML - opened

Hello, I am trying to fine-tune naver-clova-ix/donut-base with your dataset, but I'm having memory issues. Could you tell me on which machine you trained it? I am using a 4090 and it doesn't fit in memory. Your images, if I'm not mistaken, have a width of 2481 and a height of 3508, but with these numbers, I can't train it properly.

Katana ML org

Hello. I resize images to image_size = [1280, 960]. Notebook code for fine-tuning from Sparrow repo (https://github.com/katanaml/sparrow/tree/main/sparrow-ml) - https://colab.research.google.com/drive/1-v1VE2Oow_klQjO-ETGyxuIP4ebIjRGC?usp=sharing#scrollTo=PrIAfMaePYOg

Yes, indeed - Donut fine-tuning requires lots of memory. I trained on Colab Pro GPU, with 20 - 30 GB RAM.

Thank you for your reply! I'll take a look ^^

katanaml changed discussion status to closed

Sign up or log in to comment