ITM Loss Stuck at 0.63

[bfan1256]

Hi, I am trying to replicate and pretrain BLIP for distillation purposes - I am using Flickr30K + COCO and my ITM loss gets stuck at 0.63 - upon an initial look, all of the ITM predictions are 1. Is this a dataset size issue or a batch issue? I've tried changing the learning rate to a smaller LR, I've tried increasing the size of the model and more, but nothing seems to work.

[bfan1256]

It seems to predict 0s for all of the ITM labels no matter what

[ahorazahedi]

so what is solution for this problem? i am encountering same problem

[Researcher-Feng]

It is a bug of the model itself