Multiply Instead of Repeats - Option to Train on a Portion of Images Per Epoch

[a-l-e-x-d-s-9]

I'm familiar with the EveryDread2 (ED2) trainer, which provides an option to specify any portion of images from a folder. Instead of simply defining the number of repeats for a folder, you can set a 0.5 multiplier for a folder, and it will train on half of the images, selected randomly each epoch. You can also use a 1.6 multiplier or any other positive number. Negative number will discard the folder, which also important.
This approach not only includes the concept of repeats but also offers greater flexibility. It's a more effective way to balance between subjects without the need to discard any images.
I've utilized this feature in every training session with ED2, and I'm convinced that it's a crucial feature for the Kohya trainer as well.
If anyone is familiar with the Kohya codebase, could you please assess whether it's feasible to implement this feature in a reasonable manner? Your feedback would be appreciated.

[wireframewares]

Coming from EveryDream2 as well as OneTrainer, it's absolutely backwards that Kohya doesn't support this.

Being able to set a <1 decimal value in ED2 and OneTrainer is so practical and flexible, such as random image selection per epoch, as well as the sheer convenience of mixing datasets on the fly (training a person with 50 images, and mixing it with a folder full of 2,000 images of regularization ---> simply set the regularization folder to Repeats = 0.025 ... rather than having to manually remake the folder into 50 images!).