D2L-47 长短期记忆网络LSTM
Published:
本文介绍长短期记忆网络 LSTM，说明输入门、遗忘门、输出门和记忆单元如何控制序列信息，并给出 LSTM 模型的实现示例。
🎦 本节课程视频地址 👇
作者与深度学习三巨有矛盾。
长短期记忆网络
- 忘记门：将值朝0减少
- 输入门：决定不是忽略掉输入数据
- 输出门：决定是不是使用隐状态
门
\[\begin{split} &{\bf I}_t=\sigma({\bf X}_t{\bf W}_{xi}+{\bf H}_{t-1}{\bf W}_{hi}+{\bf b}_i)\\ &{\bf F}_t=\sigma({\bf X}_t{\bf W}_{xf}+{\bf H}_{t-1}{\bf W}_{hf}+{\bf b}_f)\\ &{\bf O}_t=\sigma({\bf X}_t{\bf W}_{xo}+{\bf H}_{t-1}{\bf W}_{ho}+{\bf b}_o) \end{split}\]候选记忆单元
\[{\bf \~C}_t=tanh({\bf X}_t{\bf W}_{xc}+{\bf H}_{t-1}{\bf W}_{hc}+{\bf b}_c)\]- RNN的输出
- \({\bf C, \~C}\) 与 \(\bf H\) 形状相同
记忆单元
\[{\bf C}_t={\bf F}_t⊙{\bf C}_{t-1}+{\bf I}_t⊙{\bf\~C}_t\]- \(F=0\)，舍去上一个记忆单元
- \(I=0\)，舍弃当下的记忆单元
- 两者相加，\({\bf C}_{t-1}\) 和 \({\bf\~C}_t\) 都经过非线性变换，在 \(0-1\) 之间，所以\({\bf C}_{t}\) 在 \(0-2\) 之间
- 仔细想来，\(\bf C\) 没有经过非线性，可以叠加出比较大的值
隐状态
\({\bf H}_t={\bf O}_t⊙tanh({\bf C}_t)\)
- 所以要再次非线性把 \({\bf C}_t\) clamp住。
- \({\bf O}_t=0\)，把输入和前一刻状态都舍弃，相当于重置 \({\bf H}_t\)
- \({\bf O}\) 是对 \({\bf H}\) 的保存与否
- 相当于 \({\bf C}\) 是对 \({\bf H}\) 状态的中间量
输出
\[{\bf O}_t=\phi({\bf W}_{ho}{\bf H}_t+{\bf b}_o)\]- 与RNN最后形式相同
代码实现
LSTM、GRU和RNN只是矩阵运算的forward函数上略有不同。
import torch
from torch import nn
from d2l import torch as d2l
batch_size, num_steps = 32, 35
train_iter, vocab = d2l.load_data_time_machine(batch_size, num_steps)
def get_lstm_params(vocab_size, num_hiddens, device):
num_inputs = num_outputs = vocab_size
def normal(shape):
return torch.randn(size=shape, device=device)*0.01
def three():
return (normal((num_inputs, num_hiddens)),
normal((num_hiddens, num_hiddens)),
torch.zeros(num_hiddens, device=device))
W_xi, W_hi, b_i = three() # 输入门参数
W_xf, W_hf, b_f = three() # 遗忘门参数
W_xo, W_ho, b_o = three() # 输出门参数
W_xc, W_hc, b_c = three() # 候选记忆元参数
# 输出层参数
W_hq = normal((num_hiddens, num_outputs))
b_q = torch.zeros(num_outputs, device=device)
# 附加梯度
params = [W_xi, W_hi, b_i, W_xf, W_hf, b_f, W_xo, W_ho, b_o, W_xc, W_hc,
b_c, W_hq, b_q]
for param in params:
param.requires_grad_(True)
#tensor才有requires_grad属性
return params
def init_lstm_state(batch_size, num_hiddens, device):
return (torch.zeros((batch_size, num_hiddens), device=device),#H初始化
torch.zeros((batch_size, num_hiddens), device=device)) # C初始化
def lstm(inputs, state, params):
[W_xi, W_hi, b_i, W_xf, W_hf, b_f, W_xo, W_ho, b_o, W_xc, W_hc, b_c,
W_hq, b_q] = params
(H, C) = state
outputs = []
for X in inputs:
I = torch.sigmoid((X @ W_xi) + (H @ W_hi) + b_i)
F = torch.sigmoid((X @ W_xf) + (H @ W_hf) + b_f)
O = torch.sigmoid((X @ W_xo) + (H @ W_ho) + b_o)
C_tilda = torch.tanh((X @ W_xc) + (H @ W_hc) + b_c)
C = F * C + I * C_tilda
H = O * torch.tanh(C)
Y = (H @ W_hq) + b_q
#输出还是在H上
outputs.append(Y)
return torch.cat(outputs, dim=0), (H, C)
vocab_size, num_hiddens, device = len(vocab), 256, d2l.try_gpu()
num_epochs, lr = 500, 1
model = d2l.RNNModelScratch(len(vocab), num_hiddens, device, get_lstm_params,
init_lstm_state, lstm)
d2l.train_ch8(model, train_iter, vocab, lr, num_epochs, device)
