1082ba352c5f1d524b1fcba43ee611280b169224,fairseq/trainer.py,Trainer,valid_step,#Trainer#Any#,310

Before Change

        agg_logging_output = self.criterion.__class__.aggregate_logging_outputs(logging_outputs)

        // update loss meters for validation
        if "loss" in agg_logging_output:
            self.meters["valid_loss"].update(agg_logging_output["loss"], grad_denom)
        // criterions can optionally log the NLL loss too
        if "nll_loss" in agg_logging_output:
            self.meters["valid_nll_loss"].update(agg_logging_output["nll_loss"], ntokens)

        return agg_logging_output

After Change

    def valid_step(self, sample):
        Do forward pass in evaluation mode.
        self.model.eval()

        logging_output, sample_size = {}, 0
        with torch.no_grad():
            sample = self._prepare_sample(sample)

Instances