Pytorch-LSTM

1.nn.LSTM

1.1lstm=nn.LSTM(input_size, hidden_size, num_layers)

参数：

• input_size：输入特征的维度， 一般rnn中输入的是词向量，那么 input_size 就等于一个词向量的维度，即feature_len；
• hidden_size：隐藏层神经元个数，或者也叫输出的维度（因为rnn输出为各个时间步上的隐藏状态）；
• num_layers：网络的层数；

1.2out, (h_t, c_t) = lstm(x, [h_t0, c_t0])

• x：[seq_len, batch, feature_len]
• h/c：[num_layer, batch, hidden_len]
• out：[seq_len, batch, hidden_len]

import torch
from torch import nn

lstm = nn.LSTM(input_size=100, hidden_size=20, num_layers=4)   #4层的LSTM,输入的每个词用100维向量表示,隐藏单元和记忆单元的尺寸是20

x = torch.randn(10, 3, 100)    #3句话，每句10个单词，每个单词表示为长100的向量
out, (h, c) = lstm(x)          #不传入h_0和c_0则会默认初始化
print(out.shape)               #torch.Size([10, 3, 20]) 
print(h.shape)                 #torch.Size([4, 3, 20]) 
print(c.shape)                 #torch.Size([4, 3, 20]) 

2.nn.LSTMCell

nn.LSTMCell与nn.LSTM的区别和nn.RNN与nn.RNNCell的区别一样。

2.1nn.LSTMCell()

初始化方法和上面一样。

2.2h_t, c_t = lstmcell(x_t, [h_t-1, c_t-1])

• x_t：[batch, feature_len]表示t时刻的输入
• h_t-1, c_t-1：[batch, hidden_len] t-1时刻本层的隐藏单元和记忆单元

#单层LSTM
import torch
from torch import nn

cell = nn.LSTMCell(input_size=100, hidden_size=20)       #1层的LSTM，输入的每个词用100维向量表示，隐藏单元和记忆单元的尺寸是20

h = torch.zeros(3, 20)                                   #初始化隐藏单元h和记忆单元c,取batch=3
c = torch.zeros(3, 20)

x = [torch.randn(3, 100) for _ in range(10)]             #seq_len=10个时刻的输入,每个时刻shape都是[batch,feature_len]

for xt in x:                                             #对每个时刻,传入输入xt和上个时刻的h和c
    h, c = cell(xt, (h, c))

print(h.shape,c.shape)                                   #torch.Size([3, 20]) torch.Size([3, 20])


#两层LSTM
cell_l0 = nn.LSTMCell(input_size=100, hidden_size=30)    #输入的feature_len=100,变到该层隐藏单元和记忆单元hidden_len=30
cell_l1 = nn.LSTMCell(input_size=30, hidden_size=20)     #hidden_len从l0层的30变到这一层的20

h_l0 = torch.zeros(3, 30)           #分别初始化l0层和l1层的隐藏单元h和记忆单元C,取batch=3
C_l0 = torch.zeros(3, 30)

h_l1 = torch.zeros(3, 20)
C_l1 = torch.zeros(3, 20)

x = [torch.randn(3, 100) for _ in range(10)]             #seq_len=10个时刻的输入,每个时刻shape都是[batch,feature_len]

for xt in x:
    h_l0, C_l0 = cell_l0(xt, (h_l0, C_l0))               #l0层接受xt输入
    h_l1, C_l1 = cell_l1(h_l0, (h_l1, C_l1))             #l1层接受l0层的输出h作为输入

print(h_l0.shape, C_l0.shape)                            #torch.Size([3, 30]) torch.Size([3, 30]) 
print(h_l1.shape, C_l1.shape)                            #torch.Size([3, 20]) torch.Size([3, 20])