train.py 文件源码

def compute_metrics(sess, logits_op, placeholders, data_file, exporter=None):
    """Compute metrics MAP and MRR over a dataset.

    :param sess: TensorFlow session
    :param logits_op: an operation that returns the scores for a given set of
    sentences
    :param placeholders: placeholders defined for `logits_op`
    :data_file: a HDF5 file object holding the dataset

    :returns: the values of MAP and MRR as a tuple: (MAP, MRR)
    """
    questions_ph, sentences_ph, keep_prob_ph = placeholders

    if exporter is None:
        exporter = dataio.no_op()
    next(exporter)  # priming the coroutine

    total_avep = 0.0
    total_mrr = 0.0
    n_questions = 0
    for batch in dataio.question_batches(data_file):
        feed_dict = {
            questions_ph: batch.questions,
            sentences_ph: batch.sentences,
            keep_prob_ph: 1.0
        }
        scores = logits_op.eval(session=sess, feed_dict=feed_dict)
        exporter.send(scores)

        n_questions += 1
        avep = average_precision(batch.labels, scores)
        total_avep += avep
        mrr = mean_reciprocal_rank(batch.labels, scores)
        total_mrr += mrr
    exporter.close()

    mean_avep = total_avep / n_questions
    mean_mrr = total_mrr / n_questions
    return mean_avep, mean_mrr