Yaakov Saxon
Yaakov Saxon
See [here](https://colab.research.google.com/drive/1L9zs-NH_iG8XM9zlRityAjv-uft7y5Bp?usp=sharing) for a notebook with the basic steps broken down.
Note that it is guaranteed to always return a validation set. But in various edge cases, like where there is only a single group, or if the validation fraction is...
My [first post](https://forums.fast.ai/t/developer-chat/22363/1013?u=ysaxon) on the dev forum, which this pull request was based off of, giving my justification and an example: Idea, which I already implemented for myself inside a...
@racheltho 's blog post https://www.fast.ai/2017/11/13/validation-sets/, subheading **New people, new boats, new…** is another good example.
The whole idea of GroupedSplitter is to allow fastai to automatically ensure that the pictures of the same people, same boats etc, end up wholly in either the training set...
I'll add docs and fix the merge conflicts when I get a chance. It might be a little bit, as I'm pretty busy with a few projects at the moment.
I did also already write a slightly more complicated version which will pick the most accurate (compared to requested split percentage) split out of _n_ tries (default to 5), and...
Had a couple minutes so I fixed the conflict Still need to do the rest
Will do (I haven’t forgotten about it)
Should that documentation go in 05_data.transforms.ipynb? Or somewhere else?