Function validate

language/bert/bert_code/train.py:218–283

Evaluate the model on the validation set(s) and return the losses.

(args, trainer, task, epoch_itr, subsets)

def validate(args, trainer, task, epoch_itr, subsets):
    """Evaluate the model on the validation set(s) and return the losses."""

    if args.fixed_validation_seed is not None:
        # set fixed seed for every validation
        utils.set_torch_seed(args.fixed_validation_seed)

    valid_losses = []


    for i, subset in enumerate(subsets):
        # Initialize data iterator
        itr = task.get_batch_iterator(
            dataset=task.dataset(subset),
            max_tokens=args.max_tokens_valid,
            max_sentences=args.max_sentences_valid,
            max_positions=utils.resolve_max_positions(
                task.max_positions(),
                trainer.get_model().max_positions(),
            ),
            ignore_invalid_inputs=args.skip_invalid_size_inputs_valid_test,
            required_batch_size_multiple=args.required_batch_size_multiple,
            seed=args.seed,
            num_shards=args.distributed_world_size,
            shard_id=args.distributed_rank,
            num_workers=args.num_workers,
        ).next_epoch_itr(shuffle=False)
        progress = progress_bar.build_progress_bar(
            args, itr, epoch_itr.epoch,
            prefix='valid on \'{}\' subset'.format(subset),
            no_progress_bar='simple'
        )

        epoch_loss_list = []


        # reset validation loss meters
        for k in ['valid_loss', 'valid_nll_loss']:
            meter = trainer.get_meter(k)
            if meter is not None:
                meter.reset()
        extra_meters = collections.defaultdict(lambda: AverageMeter())

        for sample in progress:
            log_output = trainer.valid_step(sample)
            epoch_loss_list += log_output['losses']
            for k, v in log_output.items():
                if k in ['loss', 'nll_loss', 'ntokens', 'nsentences', 'sample_size']:
                    continue
                extra_meters[k].update(v)

        epoch_loss_list = np.array(epoch_loss_list)
        #np.save('loss_stats/%s_valid%d_loss.npy'%(args.sess, i), epoch_loss_list)

        # log validation stats
        stats = get_valid_stats(trainer, args, extra_meters)
        for k, meter in extra_meters.items():
            stats[k] = meter.avg