Image dropout enhancement?

I've been exploring a way of training a person with limited data, in which I use more/poorer quality/AI generated photos of the person in different poses and backgrounds, then continue training with a lower learning rate and just the best photos of their face.

I was wondering if it would be possible to mark inputs, whether by file structure or file name and have the dataloader drop them from the training at the end of a certain epoch?

Proposed file structure:

train_woman
|-1_sks woman
|- Epoch 1
| |- 123.png
|- Epoch 2
| |- 456.png
|- 789.png (stays until the end)

The issue does not name files or tests. Start by tracing the training dataloader and epoch loop, then determine how input paths or names could express the proposed drop schedule. Done means a tested way to exclude marked images after selected epochs while retaining unmarked images throughout training.

Written by the indexing model from the issue text.