Kaushal Trivedi

Can you update to the latest pytorch version?

Not yet. Haven’t evaluated this yet but is on my radar.

You can use p3.8xlarge and above for parallel processing across multiple gpus

You need to set multi_gpu flag to true

Thanks. Please create the pull request. Happy to merge it.

I am assuming you only need attention weights during inference time.

makes complete sense. I am going through the fast.ai course part 2 which explains the in details. will try to incorporate callbacks soon.

Yes 0.9691836689659876 is the AUC score. You can use 0.5 as the threshold for using 0 or 1. Thats whats the accuracy_thresh metric does by default.

do you get any validation metrics?

just set the model_type parameter to distilbert and the pretrained weights name as distilbert-base-uncased