I have removed the configs we use for our internal experiments.
Some files seem edited even though nothing happened due to Black. Now, this repo passes black.
We have both epoch level loss and also batch level loss being logged. Although, the batch level loss will be logged once every 50 steps (which is the pytorch lightning default).
Yes, better solutions are possible rather than setting mask to None during inference. Perhaps a separate manifest for inference? Added it as a fixme.
This PR: