There's an issue in the evaluation of FastBert when on training stage 1.
The scorer and the way the evaluation report is logged doesn't expect to receive a list of logits (corresponding to the logits of all the ramp layers).
The model probably needs to overrides both the log_eval_report and validation_epoch_end methods of the parent class.
There's an issue in the evaluation of
FastBert
when on training stage 1.The scorer and the way the evaluation report is logged doesn't expect to receive a list of logits (corresponding to the logits of all the ramp layers). The model probably needs to overrides both the
log_eval_report
andvalidation_epoch_end
methods of the parent class.