GRU门控循环单元读书笔记

1. 概述

• 不是每个观察值都是同等重要
• 想只记住相关的观察需要
  （1）能关注的机制(更新门)
  （2）能遗忘的机制(重置门)

2. GRU关键组件

2.1 门

• 重置门 (Reset-Gate)
  $$R_t=\sigma (X_t{W}_{xr}+H_{t-1}{W}_{hr}+b_r)\text{\hspace{1em}(1)}$$
• 更新门(Update-Gate)
  $$Z_t=\sigma (X_t{W}_{xz}+H_{t-1}{W}_{hz}+b_z)\text{\hspace{1em}(2)}$$
  

2.2 候选隐藏状态${\tilde{H}}_t$

$${\tilde{H}}_t=\tanh (X_t{W}_{xh}+(R_t\odot H_{t-1})W_{hh}+b_h)\text{\hspace{1em}(3)}$$
注：因为$0\le R_t\le 1$；所以我们可以通过$R_t$来表示${\tilde{H}}_t$有数据来自于过去$H_{t-1}$

• R_t=0:表示完全丢弃过去的信息，相当于重置，故为重置门
• R_t=1:表示完全接受过去的信息，相当于RNN
  

2.3 隐状态 $H_t$

$$H_t=Z_t\odot H_{t-1}+(1-Z_t)\odot {\tilde{H}}_t\text{\hspace{1em}(4)}$$

2.4 小结

$$R_t=\sigma (X_t{W}_{xr}+H_{t-1}{W}_{hr}+b_r)$$
$$Z_t=\sigma (X_t{W}_{xz}+H_{t-1}{W}_{hz}+b_z)$$
$${\tilde{H}}_t=\tanh (X_t{W}_{xh}+(R_t\odot H_{t-1})W_{hh}+b_h)$$
$$H_t=Z_t\odot H_{t-1}+(1-Z_t)\odot {\tilde{H}}_t$$

• 重置门有助于捕获序列中的短期依赖关系
• 更新门有助于捕获序列中的长期依赖关系

3. 代码

• 源码

# -*- coding: utf-8 -*-
# @Project: zc
# @Author: zc
# @File name: GRU_test
# @Create time: 2022/1/27 16:10

# 1. 导入相关数据库
import matplotlib.pyplot as plt
import torch
from torch import nn
from d2l import torch as d2l

# 2. 定义批量大小和文本长度
batch_size, num_steps = 32, 35
train_iter, vocab = d2l.load_data_time_machine(batch_size, num_steps)


# 3. 设置GRU网络所需的参数
def get_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_xz, W_hz, b_z = three()  # 更新门参数
	W_xr, W_hr, b_r = three()  # 重置门参数
	W_xh, W_hh, b_h = three()  # 候选隐状态参数
	# 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_xz, W_hz, b_z, W_xr, W_hr, b_r, W_xh, W_hh, b_h, W_hq, b_q]
	for param in params:
		param.requires_grad_(True)
	return params


# 4. 初始化 GRU 的隐变量的状态
def init_gru_state(batch_size, num_hiddens, device):
	return (torch.zeros((batch_size, num_hiddens), device=device),)


# 5. 自定义 GRU 网络的 forward_fn;如果用到pytoch自带的，则不需要
def gru(inputs, state, params):
	# 参数展开
	W_xz, W_hz, b_z, W_xr, W_hr, b_r, W_xh, W_hh, b_h, W_hq, b_q = params
	# 初始化参数
	H, = state
	# 输出
	outputs = []
	for X in inputs:
		# update_gate 更新门
		Z = torch.sigmoid((X @ W_xz) + (H @ W_hz) + b_z)
		# reset_gate 重置门
		R = torch.sigmoid((X @ W_xr) + (H @ W_hr) + b_r)
		# 候选隐藏状态
		H_tilda = torch.tanh((X @ W_xh) + ((R * H) @ W_hh) + b_h)
		# 隐藏状态
		H = Z * H + (1 - Z) * H_tilda
		# 输出 Y
		Y = H @ W_hq + b_q
		outputs.append(Y)
	return torch.cat(outputs, dim=0), (H,)


vocab_size, num_hiddens, device = len(vocab), 256, d2l.try_gpu()
num_epochs, lr = 500, 1

num_inputs = vocab_size
gru_layer = nn.GRU(num_inputs, num_hiddens)
model = d2l.RNNModel(gru_layer, len(vocab))
model = model.to(device)
d2l.train_ch8(model, train_iter, vocab, lr, num_epochs, device)

plt.show()

• 结果

perplexity 1.0, 408368.9 tokens/sec on cuda:0
time traveller fired as we lat said ally can andither at seare y
traveller sith a slight accession ofcheerfulness really thi

4. torch.nn.GRU

class torch.nn.GRU(*args, **kwargs)

• 说明：
  将多层门控循环单元(GRU) RNN应用于输入序列。对于输入序列中的每个元素，每个层计算以下函数
  

• input_size:输入x中预期特征的数量

• hidden_size:处于隐藏状态h的特性数

• num_layers:循环层数

• 代码说明：

import torch
from torch import nn

# GRU参数来说 input_size=10,output_size=20,layers=2
# 所以输入的最后一个维度必须是10，即 input[-1]=10
# nn.GRU(input_size=10,hidden_size=20,num_layers=2)
# L = sequence_length;N=Batch_size;H_in=input_size;H_out=hidden_size
# input =(L=8,N=6,H_in=10)
# h0=(2,N=6,H_out=20)
# output = (L=8,N=6,H_out=20)
# h_n=(2,N=6,H_out=20)
rnn = nn.GRU(10, 20, 2)
input = torch.randn(8, 6, 10)
h0 = torch.randn(2, 6, 20)
output, hn = rnn(input, h0)
print(f"input.shape={input.shape}")
print(f"h0.shape={h0.shape}")
print(f"output.shape={output.shape}")
print(f"hn.shape={hn.shape}")
print(f"rnn={rnn}")

input.shape=torch.Size([8, 6, 10])
h0.shape=torch.Size([2, 6, 20])
output.shape=torch.Size([8, 6, 20])
hn.shape=torch.Size([2, 6, 20])
rnn=GRU(10, 20, num_layers=2)