MRC中答案预测实现（numpy、torch、tensorflow）

import tensorflow as tf
import numpy as np
import torch

def answer_interval_numpy(prob_s, prob_e, answer_max_len=None):
    """

    :param prob_s:  shape=[seq_len]
    :param prob_e:  shape= [seq_len]
    :param answer_max_len:  int,
    :return:
    """

    answer_probs = np.matmul(prob_s, np.transpose(prob_e, axes=-1))  # dtype: np.array
    if answer_max_len is not None:
        answer_probs = np.tril(np.triu(answer_probs, k=0), k=answer_max_len-1)

    best_start, best_end = np.unravel_index(np.argmax(answer_probs), answer_probs.shape)
    return best_start, best_end



def answer_interval_torch(prob_s, prob_e, max_answer_len=None):
    # 设文本共m个词，prob_s是大小为m的开始位置概率，prob_e是大小为m的结束位置概率，均为一维PyTorch张量
    # L为答案区间可以包含的最大的单词数
    # 输出为概率最高的区间在文本中的开始和结束位置
    # 获得m×m的矩阵，其中prob[i,j]=prob_s[i]×prob_e[j]
    prob = torch.ger(prob_s, prob_e)
    # 将prob限定为上三角矩阵，且只保留主对角线及其右上方L-1条对角线的值，其他值清零
    if max_answer_len is not None:
        # 即如果i>j或j-i+1>L，设置prob[i, j] = 0
        prob.triu_().tril_(max_answer_len - 1)
    # 转化成为numpy数组
    prob = prob.numpy()
    # 获得概率最高的答案区间，开始位置为第best_start个词, 结束位置为第best_end个词
    best_start, best_end = np.unravel_index(np.argmax(prob), prob.shape)
    return best_start, best_end


def answer_interval_tf(prob_s, prob_e, max_answer_len=None):
    """

    :param prob_s: [seq_len]
    :param prob_e:  [seq_len]
    :param max_answer_len:
    :return:
    """

    answer_probs = tf.matmul(prob_s, prob_e, transpose_b=True)
    answer_triu = tf.experimental.numpy.triu(answer_probs, k=0)
    if max_answer_len is not None:
        answer_triu = tf.experimental.numpy.tril(answer_triu, k=max_answer_len-1)

    indices = tf.unravel_index(tf.argmax(answer_triu), answer_triu.shape)
    return indices