Inconsistency in accuracy between different versions of pytorch/fairseq

Hi all,
I am posting this question here as fairseq is a part of the pytorch organization and is broadly used in nlp applications.

I have implemented a translation task using an old version of fairseq and received an accuracy of ‘x’. I then upgraded the pytorch (from 1.1 to 1.3) and fairseq versions to the latest stable versions, and then tried rerunning my training job, keeping the parameters the same. The accuracy that I received from this experiment is ‘x/3’. Can someone please help me understand what has changed in particular with the translation task?

Also please note:

the setup remains the same (multi-gpu)

the model has stopped converging like before

Please let me know if any further information is required!

It seems the issue is being tracked here.

yes. I have created that issue :slight_smile: