Few notes on the image test split of this repo.

[jamiechoi1995]

The split in this repo is different from the split on https://cs.stanford.edu/people/karpathy/deepimagesent/caption_datasets.zip. (AKA Karpathy's split in many papers. see also https://github.com/kelvinxu/arctic-captions/tree/master/splits)

In this repo, images of training and val set are first combined together, (see: https://github.com/karpathy/neuraltalk2/blob/master/coco/coco_preprocess.ipynb)

then they were shuffled and split into training, val, test set, (see:

neuraltalk2/prepro.py

Line 159 in bd8c9d8

shuffle(imgs) # shuffle the order

)

which results that in test set, there are lots of images originally comes from the training set.

But for the test split on https://cs.stanford.edu/people/karpathy/deepimagesent/caption_datasets.zip.

only contains images originally comes from the val set.

Just note this here to avoid confusion when reproducing results of papers based on Karpathy's split.