BiLSTM+ Attention Pytorch实现

最近写算法的时候发现网上关于BiLSTM加Attention的实现方式五花八门，其中很多是错的，自己基于PyTorch框架实现了一版，主要用到了LSTM处理变长序列和masked softmax两个技巧。代码如下：

1、attention_utils.py

from typing import Dict, Optional

import numpy as np
import torch
import torch.nn.functional as F
from torch import Tensor


def create_src_lengths_mask(
    batch_size: int, src_lengths: Tensor, max_src_len: Optional[int] = None
):
    """
    Generate boolean mask to prevent attention beyond the end of source
    Inputs:
      batch_size : int
      src_lengths : [batch_size] of sentence lengths
      max_src_len: Optionally override max_src_len for the mask
    Outputs:
      [batch_size, max_src_len]
    """
    if max_src_len is None:
        max_src_len = int(src_lengths.max())
    src_indices = torch.arange(0, max_src_len).unsqueeze(0).type_as(src_lengths)
    src_indices = src_indices.expand(batch_size, max_src_len)
    src_lengths = src_lengths.unsqueeze(dim=1).expand(batch_size, max_src_len)
    # returns [batch_size, max_seq_len]
    return (src_indices < src_lengths).int().detach()


def masked_softmax(scores, src_lengths, src_length_masking=True):
    """Apply source length masking then softmax.
    Input and output have shape bsz x src_len"""
    if src_length_masking:
        bsz, max_src_len = scores.size()
        # compute masks
        src_mask = create_src_lengths_mask(bsz, src_lengths)
        # Fill pad positions with -inf
        scores = scores.masked_fill(src_mask == 0, -np.inf)

    # Cast to float and then back again to prevent loss explosion under fp16.
    return F.softmax(scores.float(), dim=-1).type_as(scores)

2、layers.py

import torch
from torch import nn
import torch.nn.functional as F

from utils.attention_utils import masked_softmax


# s(x, q) = v.T * tanh (W * x + b)
class MLPAttentionNetwork(nn.Module):

    def __init__(self, hidden_dim, attention_dim, src_length_masking=True):
        super(MLPAttentionNetwork, self).__init__()

        self.hidden_dim = hidden_dim
        self.attention_dim = attention_dim
        self.src_length_masking = src_length_masking

        # W * x + b
        self.proj_w = nn.Linear(self.hidden_dim, self.attention_dim, bias=True)
        # v.T
        self.proj_v = nn.Linear(self.attention_dim, 1, bias=False)

    def forward(self, x, x_lengths):
        """
        :param x: seq_len * batch_size * hidden_dim
        :param x_lengths: batch_size
        :return: batch_size * seq_len, batch_size * hidden_dim
        """
        seq_len, batch_size, _ = x.size()
        # (seq_len * batch_size, hidden_dim)
        # flat_inputs = x.view(-1, self.hidden_dim)
        flat_inputs = x.reshape(-1, self.hidden_dim)
        # (seq_len * batch_size, attention_dim)
        mlp_x = self.proj_w(flat_inputs)
        # (batch_size, seq_len)
        att_scores = self.proj_v(mlp_x).view(seq_len, batch_size).t()
        # (seq_len, batch_size)
        normalized_masked_att_scores = masked_softmax(
            att_scores, x_lengths, self.src_length_masking
        ).t()
        # (batch_size, hidden_dim)
        attn_x = (x * normalized_masked_att_scores.unsqueeze(2)).sum(0)

        return normalized_masked_att_scores.t(), attn_x

3、model.py

import torch
import torch.nn as nn
from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence

from layers import MLPAttentionNetwork


class BiLSTMAttentionNetwork(nn.Module):

    def __init__(self, num_vocab, embedding_dim, max_len, hidden_dim, num_layers, bidirectional, attention_dim, num_classes):

        super(BiLSTMAttentionNetwork, self).__init__()

        # 词表长度，实际单词数量+填充占位符（1）
        self.num_vocab = num_vocab
        # 最大序列长度
        self.max_len = max_len
        # 单词隐向量维度
        self.embedding_dim = embedding_dim
        # LSTM中隐层的维度
        self.hidden_dim = hidden_dim
        # 循环神经网络的层数
        self.num_layers = num_layers
        # 是否使用双向RNN，布尔值
        self.bidirectional = bidirectional
        # 注意力层参数维度
        self.attention_dim = attention_dim
        # 标签数量
        self.num_classes = num_classes
        # Embedding层
        self.embedding_layer = nn.Embedding(self.num_vocab, self.embedding_dim, padding_idx=0)
        # RNN层
        self.bilstm_layer = nn.LSTM(self.embedding_dim, self.hidden_dim, self.num_layers, bidirectional=self.bidirectional,
                              batch_first=True)
        # MLP注意力层
        self.mlp_attention_layer = MLPAttentionNetwork(2 * self.hidden_dim, self.attention_dim)
        # 全连接层
        self.fc_layer = nn.Linear(2 * self.hidden_dim, self.num_classes)
        # 单层softmax分类器
        self.softmax_layer = nn.Softmax(dim=1)

    def forward(self, x, lengths):
        """
        :param x: 填充好的序列
        :param lengths:
        :return:
        """

        # x: t.tensor([[1,2,3],[6,0,0],[4,5,0], [3, 7, 1]])
        # lengths: t.tensor([3, 1, 2, 3])、序列的实际长度

        x_input = self.embedding_layer(x)
        # print(x_input)
        x_packed_input = pack_padded_sequence(input=x_input, lengths=lengths, batch_first=True, enforce_sorted=False)
        # print(x_packed_input)
        packed_out, _ = self.bilstm_layer(x_packed_input)
        # print(packed_out)
        outputs, _ = pad_packed_sequence(packed_out, batch_first=True, total_length=self.max_len, padding_value=0.0)
        # print(out)
        atten_scores, atten_out = self.mlp_attention_layer(outputs.permute(1, 0, 2), lengths)
        # print(atten_out)
        # (batch_size, num_classes)
        logits = self.softmax_layer(self.fc_layer(atten_out))
        return atten_scores, logits

测试代码如下：

if __name__ == '__main__':
    # num_vocab, embedding_dim, hidden_dim, num_layers, bidirectional, attention_dim, num_classes
    b_a = BiLSTMAttentionNetwork(20, 3, 3, 2, 1, bidirectional=True, attention_dim=5, num_classes=10)
    x = torch.tensor([[1, 2, 3], [6, 0, 0], [4, 5, 0], [3, 7, 1]])
    lengths = torch.tensor([3, 1, 2, 3])
    atten_scores, logits = b_a(x, lengths)
    print('--------------------->注意力分布')
    print(atten_scores)
    print('--------------------->预测概率')
    print(logits)

效果如下：

--------------------->注意力分布:
tensor([[0.3346, 0.3320, 0.3335],
        [1.0000, 0.0000, 0.0000],
        [0.4906, 0.5094, 0.0000],
        [0.3380, 0.3289, 0.3330]], grad_fn=<TBackward>)
--------------------->预测概率:
tensor([[0.0636, 0.0723, 0.1241, 0.0663, 0.0671, 0.0912, 0.1244, 0.1446, 0.1223,
         0.1240],
        [0.0591, 0.0745, 0.1264, 0.0650, 0.0657, 0.0853, 0.1273, 0.1478, 0.1186,
         0.1303],
        [0.0634, 0.0727, 0.1178, 0.0678, 0.0688, 0.0925, 0.1203, 0.1492, 0.1228,
         0.1249],
        [0.0615, 0.0739, 0.1253, 0.0633, 0.0675, 0.0872, 0.1226, 0.1490, 0.1224,
         0.1274]], grad_fn=<SoftmaxBackward>)