pytorch-LSTM()简单介绍

参考：https://www.cnblogs.com/duye/p/9913386.html

torch.nn.LSTM()的参数：

• input_size: 输入特征维数
• hidden_size: 隐层状态的维数
• num_layers: 层数
• bias: 隐层状态是否带bias，默认为true
• batch_first: 是否输入输出的第一维为batch_size，因为pytorch中batch_size维度默认是第二维度，通常设为True
• dropout: 是否在除最后一个RNN层外的RNN层后面加dropout层
• bidirectional: 是否是双向RNN，默认为false，若为true，则num_directions=2，否则为1

LSTM输出：

• out, hidden = lstm(input, hidden)
• out的维度：(batch_size, seq_len, hidden_size)
• hidden=(hn, cn)
• hn和cn的维度：(num_layers*num_directions, batch_size, hidden_size)

通常定义一个1层双向LSTM模型如下：（NER任务，输出是对每个字分类）

import torch
import torch.nn as nn


class MyBiLSTM(nn.Module):
    def __init__(self, vocab_size, embedding_dim, hidden_dim, tag_size):
        super(MyBiLSTM, self).__init__()
        self.vocab_size = vocab_size
        self.embedding_dim = embedding_dim
        self.hidden_dim = hidden_dim
        self.tag_size = tag_size
        self.embedding_layer = nn.Embedding(vocab_size, embedding_dim)
        self.lstm_layer = nn.LSTM(embedding_dim, hidden_dim, num_layers=1, bidirectional=True, batch_first=True)
        # *2因为是双向的
        self.output_layer = nn.Linear(hidden_dim * 2, tag_size)

    def init_hidden(self, batch_size):
        # 正太分布, num_layer=2, 因为是双向的
        return (torch.randn(2, batch_size, self.hidden_dim), torch.randn(2, batch_size, self.hidden_dim))

    def forward(self, input):
        self.hidden = self.init_hidden(len(input))
        embeds = self.embedding_layer(input)
        lstm_out, self.hidden = self.lstm_layer(embeds, self.hidden)
        model_out = self.output_layer(lstm_out)
        return model_out

model = MyBiLSTM(5, 10, 10, 4)  # 5x10x10x4的LSTM
input = torch.tensor([[0, 1, 2, 3, 4]]) # 输入5个字符
print(model(input)) # 输出5x4的矩阵