Gradual image size increase

fastai
1

This article mention gradually increasing image size during training. https://www.fast.ai/2018/08/10/fastai-diu-imagenet/ You need to be using adaptive pooling to take into account the different image sizes…

But I am wondering how this is implemented in terms of image resizing scheduling. Do they first fit one epoch with a specific size, then increase the size a little bit for the next epoch like it is shown in the courses, or are they using something more clever like gradually increasing the image size while fitting one epoch? I would be interested in finding more information about this and how this could be implemented in fastai.

Thanks!

2

I’m currently doing an experiment for this somewhat with transfer learning. I’m hoping to have a medium article or at the very least a fastai post on this here within the next week.

1 Like
3

Excellent could you update this post once you finish your article?

Thanks!

1 Like
4

Absolutely!