def reduce_metrics(logging_outputs) -> None:
"""Aggregate logging outputs from data parallel training."""
CrossEntropyCriterion.reduce_metrics(logging_outputs)
num_corr = sum(log.get("num_corr", 0) for log in logging_outputs)
num_tot = sum(log.get("num_tot", 0) for log in logging_outputs)
metrics.log_scalar("accuracy",
num_corr.float() / num_tot *
100 if num_tot > 0 else 0.0,
num_tot,
round=3)
def reduce_metrics(logging_outputs) -> None:
"""Aggregate logging outputs from data parallel training."""
CrossEntropyCriterion.reduce_metrics(logging_outputs)
word_error = sum(log.get('word_error', 0) for log in logging_outputs)
word_count = sum(log.get('word_count', 0) for log in logging_outputs)
char_error = sum(log.get('char_error', 0) for log in logging_outputs)
char_count = sum(log.get('char_count', 0) for log in logging_outputs)
if word_count > 0: # model.training == False
metrics.log_scalar('wer', float(word_error) / word_count * 100, word_count, round=4)
if char_count > 0: # model.training == False
metrics.log_scalar('wer', float(word_error) / word_count * 100, word_count, round=4)