Flowers-102 train split include validation split ?

[hankyul2]

Hi! I am wondering if train size (2040) of flowers-102 in table 5 (see below) means that we should use both train split(1020) and validation split(1020) in training flowers-102 dataset ?

This table is from ResNet strikes back: An improved training procedure in timm, 2021, Wightman◦

Thank you.

[rwightman]

@hankyul2 yes, it seems fairly common in a few papers that reference Flowers-102 as having '2040' train that the default train + val split are combined and the 6149 'test' is used for test or val (depending how the train / hparam selection is done)

[hankyul2]

Thank you for your answer. It helps me a lot.

[laimleon]

It’s interesting how the Flowers‑102 dataset is split for training and validation. It reminds me of tipos de flores in real life — just like some flowers bloom in small batches to test the environment before planting a full garden, dataset splits allow models to “learn” safely from part of the data before being evaluated on the full set. Understanding both helps achieve the best results.