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

Before Change

            logging_outputs = [logging_output]

        // aggregate stats and logging outputs
        ntokens = sum(log.get("ntokens", 0) for log in logging_outputs)
        grad_denom = self.criterion.__class__.grad_denom(sample_sizes)
        agg_logging_output = self.criterion.__class__.aggregate_logging_outputs(logging_outputs)

After Change

        self.model.eval()

        logging_output, sample_size = {}, 0
        with torch.no_grad():
            sample = self._prepare_sample(sample)
            if sample is None:
                sample = self._prepare_sample(self._dummy_batch)
            _loss, sample_size, logging_output = self.task.get_loss(
                self.model, self.criterion, sample,
            )

        // gather logging outputs from all replicas
        if self.args.distributed_world_size > 1:
            logging_output, sample_size = zip(*distributed_utils.all_gather_list(
                [logging_output, sample_size],
            ))

Instances