def aggregate_logging_outputs()

in custom/sequence_penalty_loss.py [0:0]

• 20 lines of code
• 11 McCabe index (conditional complexity)

    def aggregate_logging_outputs(logging_outputs):
        """Aggregate logging outputs from data parallel training."""
        loss_sum = sum(log.get('seq_loss', 0) for log in logging_outputs)
        ntokens = sum(log.get('seq_ntokens', 0) for log in logging_outputs)
        nsentences = sum(log.get('seq_nsentences', 0) for log in logging_outputs)
        sample_size = sum(log.get('seq_sample_size', 0) for log in logging_outputs)
        repeat_mask = sum(log.get('seq_repeat_mask', 0) for log in logging_outputs)

        agg_output = {
            'seq_loss': loss_sum / max(sample_size, 1.0) / math.log(2),
            'seq_ntokens': ntokens,
            'seq_nsentences': nsentences,
            'seq_sample_size': sample_size,
            'seq_repeat_mask': repeat_mask / sample_size if sample_size > 0 else 0
        }

        for n in range(1, 5):
            key = 'pct_repeat_%dgrams' % n
            ngram_repeats = sum(log.get(key, 0) for log in logging_outputs)
            # Normalize by the number of sentences since this is the sum of per-sentence metrics.
            agg_output['seq_' + key] = ngram_repeats /nsentences if nsentences > 0 else 0

        if sample_size != ntokens:
            agg_output['seq_nll_loss'] = loss_sum / max(ntokens, 1.0) / math.log(2)
        return agg_output