Train/val loss on flickr30k

[yjsong22]

Hi,
Thanks for your nice work. I want to ask what are the training and validation losses for flickr 30k dataset in your experiments. I experimented with your code on flickr 30k but it seems to be overfitting and the losses are very high (about 10.7). Do you have any ideas on this?
Thank you.

[jmisilo]

Hey, so the train loss was 10.03 and the validation loss was 10.02. These are quite high, but the challenging part was reducing them + computing power was limited. I don't have any good ideas now how to improve this, maybe playing with forward_expansion and num_layers layers of Mapping?