关注每一个序列

1.不是每个观察值都是同等重要

2.想只记住的观察需要：能关注的机制（更新门 update gate）、能遗忘的机制（重置门 reset gate）

python">!pip install --upgrade d2l==0.17.5  #d2l需要更新

python">import torch
from torch import nn
from d2l import torch as d2lbatch_size, num_steps = 32, 35
train_iter, vocab = d2l.load_data_time_machine(batch_size, num_steps)

Downloading ../data/timemachine.txt from http://d2l-data.s3-accelerate.amazonaws.com/timemachine.txt...

下一步是初始化模型参数。 我们从标准差为0.01的高斯分布中提取权重， 并将偏置项设为0，超参数num_hiddens定义隐藏单元的数量， 实例化与更新门、重置门、候选隐状态和输出层相关的所有权重和偏置。

python">def get_params(vocab_size, num_hiddens, device):num_inputs = num_outputs = vocab_sizedef normal(shape):return torch.randn(size=shape, device=device)*0.01def 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

将定义隐状态的初始化函数init_gru_state。此函数返回一个形状为（批量大小，隐藏单元个数）的张量，张量的值全部为零。

python">def init_gru_state(batch_size, num_hiddens, device):return (torch.zeros((batch_size, num_hiddens), device=device), )

准备定义门控循环单元模型， 模型的架构与基本的循环神经网络单元是相同的， 只是权重更新公式更为复杂。

python">def gru(inputs, state, params):W_xz, W_hz, b_z, W_xr, W_hr, b_r, W_xh, W_hh, b_h, W_hq, b_q = paramsH, = stateoutputs = []for X in inputs:Z = torch.sigmoid((X @ W_xz) + (H @ W_hz) + b_z)R = torch.sigmoid((X @ W_xr) + (H @ W_hr) + b_r)H_tilda = torch.tanh((X @ W_xh) + ((R * H) @ W_hh) + b_h)H = Z * H + (1 - Z) * H_tildaY = H @ W_hq + b_qoutputs.append(Y)return torch.cat(outputs, dim=0), (H,)

训练结束后，我们分别打印输出训练集的困惑度， 以及前缀“time traveler”和“traveler”的预测序列上的困惑度。

python">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_params,init_gru_state, gru)
d2l.train_ch8(model, train_iter, vocab, lr, num_epochs, device)

perplexity 1.1, 31831.9 tokens/sec on cuda:0
time traveller for so it will be convenient to speak of himwas e
travelleryou can show black is white by argument said filby

简洁实现

python">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)

perplexity 1.0, 255484.2 tokens/sec on cuda:0
time traveller for so it will be convenient to speak of himwas e
traveller with a slight accession ofcheerfulness really thi