When evaluating, shouldn't we always use the SequentialSampler instead of DistributedSampler? Evaluation only runs on 1 GPU no matter what, so if you use the DistributedSampler with N GPUs, I think you'll only evaluate on 1/N of the evaluation set. That's at least what I'm finding when I run an older/modified version of this repo.
Currently the L2 regularization is hard-coded to "0.01", even though there is a --weight_decay flag implemented (that is unused). I'm making this flag control the weight decay used for fine-tuning in this script.
Ethan Perez