You can’t do that directly since the training dataloader shuffles the data by default (so the images are not the first items). If you use the validation dataloader, it loads the image in order, so you can find their filenames in data.valid_ds.items[0] (or 1, 2, …)
4 Likes