传统RNN网络及其案例--人名分类

传统的RNN模型简介

RNN

先上图

这图看起来莫名其妙，想拿着跟CNN对比着学第一眼看上去有点摸不着头脑，其实我们可以把每一个时刻的图展开来，如下

其中，为了简化计算，我们默认每一个隐层参数相同，这样看来RNN的结构就比较简单了，相比较CNN来说，RNN引入了更多的时序信息。

LSTM

在训练 RNN 时，每个时间步的输出都依赖于之前时间步的状态，这种依赖关系形成了一个链式结构。当反向传播时，梯度需要通过多个时间步传播回去，由于链式法则的存在，这个过程中梯度会多次进行乘法运算。如果这些乘法运算的结果小于1，梯度就会随着时间步的增加逐渐衰减，最终可能消失到几乎为零，就会导致梯度消失。RNN 中常用的激活函数如 Sigmoid 或者 tanh 函数，它们的输出范围都在 (0, 1) 或者 (-1, 1) 之间。在反向传播时，如果梯度在这些函数的导数范围内，则可以稳定地传播；但如果超出了这个范围，梯度可能会指数级增长或减少，导致梯度爆炸。而且这些在处理长序列时特别容易发生，因此，出现了RNN的改良版，LSTM。

先看图：

谈到LSTM就无法避免的提及它的三个门和最上面的记忆单元

记忆单元：记忆单元是LSTM的核心，用于存储信息。它可以看作是一条信息通道，贯穿整个 LSTM单元链条，允许信息直接传递，减少信息丢失。
遗忘门：遗忘门决定哪些信息需要从记忆单元中删除。它通过sigmoid函数（将当前输入和前一时刻的隐藏状态作为输入）输出一个0到1之间的值。接近0的值表示需要遗忘的信息，接近1的值表示需要保留的信息。

\[f_t = \sigma(W_f \cdot [h_{t-1}, x_t] + b_f)
\]
输入门：输入门决定哪些新的信息需要添加到记忆单元中。它由两个部分组成：一个sigmoid层决定哪些值将被更新；一个tanh层生成新的候选值向量。

\[i_t = \sigma(W_i \cdot [h_{t-1}, x_t] + b_i)
\]

\[\tilde{C}_t = \tanh(W_C \cdot [h_{t-1}, x_t] + b_C)
\]
输出门：输出门决定记忆单元的哪些部分将被输出作为当前时刻的隐藏状态。它通过sigmoid层和tanh层来实现。

\[o_t = \sigma(W_o \cdot [h_{t-1}, x_t] + b_o)
\]

\[h_t = o_t * \tanh(C_t)
\]

LSTM的工作流程如下

遗忘阶段：计算遗忘门的值，以确定当前记忆单元状态中需要遗忘的部分。

\[C_t = f_t * C_{t-1}
\]
输入阶段：计算输入门的值，并生成新的候选记忆内容。

\[C_t = C_t + i_t * \tilde{C}_t
\]
更新记忆单元：结合遗忘门和输入门的输出，更新当前记忆单元的状态。
输出阶段：计算输出门的值，并生成新的隐藏状态。

完整公式流程：

\[f_t = \sigma(W_f \cdot [h_{t-1}, x_t] + b_f)
\]

\[i_t = \sigma(W_i \cdot [h_{t-1}, x_t] + b_i)
\]

\[\tilde{C}_t = \tanh(W_C \cdot [h_{t-1}, x_t] + b_C)
\]

\[C_t = f_t * C_{t-1} + i_t * \tilde{C}_t
\]

\[o_t = \sigma(W_o \cdot [h_{t-1}, x_t] + b_o)
\]

\[h_t = o_t * \tanh(C_t)
\]

GRU

LSTM固然很强，解决了RNN对于长序列模型表现很拉跨的难题，但是仔细查看LSTM的过程就会发现，相对于RNN来说他引入了太多的参数，很容易就过拟合和训练时间大大加长，因此，GRU改进这一问题

更新门：更新门控制着前一时间步的信息和当前时间步的新信息之间的混合。它通过sigmoid函数决定有多少过去的信息需要保留，以及有多少新的信息需要添加。

\[z_t = \sigma(W_z \cdot [h_{t-1}, x_t] + b_z)
\]
重置门：重置门控制着前一时间步的隐藏状态在当前时间步中被遗忘的比例。它通过sigmoid函数决定有多少前一时间步的信息需要被重置。

\[r_t = \sigma(W_r \cdot [h_{t-1}, x_t] + b_r)
\]
候选隐藏状态：候选隐藏状态结合了重置门的结果，决定当前时间步的隐藏状态。

\[\tilde{h}_t = \tanh(W \cdot [r_t * h_{t-1}, x_t] + b)
\]
隐藏状态：最终的隐藏状态是更新门和候选隐藏状态的组合。

\[h_t = (1 - z_t) * h_{t-1} + z_t * \tilde{h}_t
\]

工作流程如下：

重置阶段：计算重置门的值，以确定前一时间步的信息在当前时间步中被重置的比例。

\[r_t = \sigma(W_r \cdot [h_{t-1}, x_t] + b_r)
\]
更新阶段：计算更新门的值，以确定有多少信息从前一时间步保留到当前时间步。

\[z_t = \sigma(W_z \cdot [h_{t-1}, x_t] + b_z)
\]
候选隐藏状态阶段：计算候选隐藏状态，该状态结合了重置门的结果和当前输入信息。

\[\tilde{h}_t = \tanh(W \cdot [r_t * h_{t-1}, x_t] + b)
\]
隐藏状态更新阶段：结合更新门和候选隐藏状态，更新当前时间步的隐藏状态。

\[h_t = (1 - z_t) * h_{t-1} + z_t * \tilde{h}_t
\]

完整工作流程

\[z_t = \sigma(W_z \cdot [h_{t-1}, x_t] + b_z)
\]

\[r_t = \sigma(W_r \cdot [h_{t-1}, x_t] + b_r)
\]

\[\tilde{h}_t = \tanh(W \cdot [r_t * h_{t-1}, x_t] + b)
\]

\[h_t = (1 - z_t) * h_{t-1} + z_t * \tilde{h}_t
\]

使用传统RNN模型来进行人名分类

1. 准备工作

要用到的数据集点此下载，备用地址，点击下载

导入一些包和写一个读取数据的函数（这段代码不是重点，直接抄就行了，只需要记住几个关键的变量）

category_lines: 人名类别与具体人名对应关系的字典，形式为{人名类别:[人名1，人名2，...]}

all_categories：所有的类别构成的列表

all_letters：所有的字符

import torch

from io import open

import glob

import os

import unicodedata

import string

import random

import time

import math

import torch.nn as nn

import matplotlib.pyplot as plt

data_path = './data/names/'

all_letters = string.ascii_letters + " .,;'"

def unicodeToAscii(text):

    """

    Converts a Unicode string to an ASCII string.

    Args:

        text (str): The Unicode string to convert.

    Returns:

        str: The ASCII string.

    """

    return ''.join([

        unicodedata.normalize('NFKD', char)

        for char in text

        if not unicodedata.combining(char)

    ]).encode('ascii', 'ignore').decode('ascii')

def readLines(filename):

    lines = open(filename, encoding='utf-8').read().strip().split('\n')

    return [unicodeToAscii(line) for line in lines]

# 构建一个人名类别与具体人名对应关系的字典

category_lines = {}

# 构建所有类别的列表

all_categories = []

# 遍历所有的文件，使用glob.glob中可以利用正则表达式遍历

for filename in glob.glob(data_path + '*.txt'):

    category = os.path.splitext(os.path.basename(filename))[0]

    all_categories.append(category)

    lines = readLines(filename)

    # 将类别与人名对应关系存储到字典中

    category_lines[category] = lines

# 测试

print(category_lines['Italian'][:5])

字符无法直接被网络识别，因此要将其编码，这里使用最简单的one-hot编码，实现一个函数lineToTensor(line)，将输入的名字编码成张量

def lineToTensor(line):

    # 首先初始化一个全0的张量，大小为len(line) * 1 * n_letters

    # 代表人名每个字母都用一个(1 * n_letters)的one-hot向量表示

    tensor = torch.zeros(len(line), 1, len(all_letters))

    # 遍历人名的每个字母, 并搜索其在所有字母中的位置，将其对应的位置置为1

    for li, letter in enumerate(line):

        tensor[li][0][all_letters.find(letter)] = 1

    return tensor

# 测试

line = "Bai"

tensor = lineToTensor(line)

print("line_tensor:", tensor)

print("line_tensor_size:", tensor.size())

2. 模型搭建

搭建RNN模型

class RNN(nn.Module):

    def __init__(self, input_size, hidden_size, output_size, num_layers=1):

        super(RNN, self).__init__()

        # input_size: 输入数据的特征维度

        # hidden_size: RNN隐藏层的最后一个维度

        # output_size: RNN网络最后线性层的输出维度

        # num_layers: RNN网络的层数

        self.input_size = input_size

        self.hidden_size = hidden_size

        self.output_size = output_size

        self.num_layers = num_layers

        self.rnn = nn.RNN(input_size, hidden_size, num_layers)

        self.linear = nn.Linear(hidden_size, output_size)

        self.softmax = nn.LogSoftmax(dim=-1)

    def forward(self, input1, hidden):

        # input: 人名分类器中的输入张量，形状是1*n_letters

        # hidden: 代表隐藏层张量，形状是self.num_layers*1*hidden_size

        # 输入到RNN中的张量要求是三维张量，所以需要用unsqueeze()函数扩充维度

        input1 = input1.unsqueeze(0)

        rr, hn = self.rnn(input1, hidden)

        # 将RNN中获得的结果通过线性层变换和softmax函数输出

        return self.softmax(self.linear(rr)), hn

    def initHidden(self):

        # 初始化隐藏层张量，全0张量，维度是3

        return torch.zeros(self.num_layers, 1, self.hidden_size)

搭建LSTM模型

class LSTM(nn.Module):

    def __init__(self, input_size, hidden_size, output_size, num_layers=1):

        super(LSTM, self).__init__()

        self.input_size = input_size

        self.hidden_size = hidden_size

        self.output_size = output_size

        self.num_layers = num_layers

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

        self.linear = nn.Linear(hidden_size, output_size)

        self.softmax = nn.LogSoftmax(dim=-1)

    def forward(self, input1, hidden, c):

        # 注意：LSTM网络的输入有三个张量，不能忘记细胞状态C

        input1 = input1.unsqueeze(0)

        rr, (hn, cn) = self.lstm(input1, (hidden, c))

        return self.softmax(self.linear(rr)), hn, cn

    def initHiddenAndC(self):

        c = hidden = torch.zeros(self.num_layers, 1, self.hidden_size)

        return hidden, c

搭建GRU模型

class GRU(nn.Module):

    def __init__(self, input_size, hidden_size, output_size, num_layers=1):

        super(GRU, self).__init__()

        self.input_size = input_size

        self.hidden_size = hidden_size

        self.output_size = output_size

        self.num_layers = num_layers

        self.gru = nn.GRU(input_size, hidden_size, num_layers)

        self.linear = nn.Linear(hidden_size, output_size)

        self.softmax = nn.LogSoftmax(dim=-1)

    def forward(self, input1, hidden):

        input1 = input1.unsqueeze(0)

        rr, hn = self.gru(input1, hidden)

        return self.softmax(self.linear(rr)), hn

    def initHidden(self):

        return torch.zeros(self.num_layers, 1, self.hidden_size)

3. 模型的实例化与训练

定义一些参数以及实例化模型

# 参数

input_size = len(all_letters)

n_hidden = 128

output_size = n_categories

input1 = lineToTensor('B').squeeze(0)

hidden = c = torch.zeros(1, 1, n_hidden)

rnn = RNN(input_size, n_hidden, output_size)

lstm = LSTM(input_size, n_hidden, output_size)

gru = GRU(input_size, n_hidden, output_size)

# 测试

rnn_output, rnn_hidden = rnn(input1, hidden)

lstm_output, lstm_hidden, next_c = lstm(input1, hidden, c)

gru_output, gru_hidden = gru(input1, hidden)

# 打印输出信息

print("rnn_output:", rnn_output)

print("rnn_shape:", rnn_output.shape)

print("********************************")

print("lstm_output:", lstm_output)

print("lstm_shape:", lstm_output.shape)

print("********************************")

print("gru_output:", gru_output)

print("gru_shape:", gru_output.shape)

categoryFromOutput(output)函数功能为从模型输出中获取最大值和最大值的索引

def categoryFromOutput(output):

    # 从输出中获取最大值和最大值的索引

    top_n, top_i = output.topk(1)

    category_i = top_i[0].item()

    return all_categories[category_i], category_i

randomTrainingExample()函数功能为随机选训练所需的数据

def randomTrainingExample():

    # 随机选择一个类别

    category = random.choice(all_categories)

    # 从该类别中随机选择一个人名

    line = random.choice(category_lines[category])

    # 将人名转换为张量

    category_tensor = torch.tensor([all_categories.index(category)], dtype=torch.long)

    line_tensor = lineToTensor(line)

    return category, line, category_tensor, line_tensor

构建训练函数

# 构建传统RNN训练函数

criterion = nn.NLLLoss()

learning_rate = 0.005

def trainRNN(category_tensor, line_tensor):

    hidden = rnn.initHidden()

    rnn.zero_grad()

    output = None

    for i in range(line_tensor.size()[0]):

        output, hidden = rnn(line_tensor[i], hidden)

    # rnn对象由nn.RNN实例化得到，最终输出得到的是三维张量，为了满足category_tensor的维度要求，需要将其转换为二维张量

    loss = criterion(output.squeeze(0), category_tensor)

    loss.backward()

    # 更新参数

    for p in rnn.parameters():

        p.data.add_(-learning_rate, p.grad.data)

    return output, loss.item()

def trainLSTM(category_tensor, line_tensor):

    hidden, c = lstm.initHiddenAndC()

    lstm.zero_grad()

    output = None

    for i in range(line_tensor.size()[0]):

        output, hidden, c = lstm(line_tensor[i], hidden, c)

    loss = criterion(output.squeeze(0), category_tensor)

    loss.backward()

    for p in lstm.parameters():

        p.data.add_(-learning_rate, p.grad.data)

    return output, loss.item()

def trainGRU(category_tensor, line_tensor):

    hidden = gru.initHidden()

    gru.zero_grad()

    output = None

    for i in range(line_tensor.size()[0]):

        output, hidden = gru(line_tensor[i], hidden)

    loss = criterion(output.squeeze(0), category_tensor)

    loss.backward()

    for p in gru.parameters():

        p.data.add_(-learning_rate, p.grad.data)

    return output, loss.item()

绘图的辅助函数timeSince(since)用于记录代码运行时间

# 构建时间计算函数

def timeSince(since):

    now = time.time()

    s = now - since

    m = math.floor(s / 60)

    s -= m * 60

    return "%dm %ds" % (m, s)

构建完整的训练函数

# 训练轮次

n_iters = 1000

# 每隔print_every打印一次信息

print_every = 50

# 每个plot_every作为一次绘图采样点

plot_every = 10

def train(train_type_fn):

    # 每个制图间隔损失保存列表

    all_losses = []

    # 获得开始的时间戳

    start = time.time()

    current_loss = 0

    for iter in range(1, n_iters + 1):

        category, line, category_tensor, line_tensor = randomTrainingExample()

        output, loss = train_type_fn(category_tensor, line_tensor)

        current_loss += loss

        if iter % print_every == 0:

            guess, guess_i = categoryFromOutput(output)

            correct = "✓" if guess == category else "✗ (%s)" % category

            print("%d %d%% (%s) %.4f %s / %s %s" % (iter, iter / n_iters * 100, timeSince(start), loss, line, guess, correct))

        if iter % plot_every == 0:

            all_losses.append(current_loss / plot_every)

            current_loss = 0

    return all_losses, int(time.time() - start)

训练并绘图

# 训练并制作对比图

all_losses1, period1 = train(trainRNN)

all_losses2, period2 = train(trainLSTM)

all_losses3, period3 = train(trainGRU)

plt.figure(0)

plt.plot(all_losses1, label='RNN')

plt.plot(all_losses2, color='red', label='LSTM')

plt.plot(all_losses3, color='green', label='GRU')

plt.legend(loc='upper left')

plt.figure(1)

x_data = ['RNN', 'LSTM', 'GRU']

y_data = [period1, period2, period3]

plt.bar(range(len(x_data)), y_data, color='green', tick_label=x_data)

为方便测试，训练轮次只设置了一千，图形跑出来看不是很清楚，以下为训练1e5次的效果

4. 构建评估函数和预测函数

评估函数

def evaluateRNN(line_tensor):

    output = None

    hidden = rnn.initHidden()

    for i in range(line_tensor.size()[0]):

        output, hidden = rnn(line_tensor[i], hidden)

    return output.squeeze(0)

def evaluateLSTM(line_tensor):

    output = None

    hidden, c = lstm.initHiddenAndC()

    for i in range(line_tensor.size()[0]):

        output, hidden, c = lstm(line_tensor[i], hidden, c)

    return output.squeeze(0)

def evaluateGRU(line_tensor):

    output = None

    hidden = gru.initHidden()

    for i in range(line_tensor.size()[0]):

        output, hidden = gru(line_tensor[i], hidden)

    return output.squeeze(0)

预测函数

def predict(input_line, evaluate, n_predictions=3):

    print("\n> %s" % input_line)

    with torch.no_grad():

        output = evaluate(lineToTensor(input_line))

        topv, topi = output.topk(n_predictions, 1, True)

        predictions = []

        for i in range(n_predictions):

            value = topv[0][i].item()

            category_index = topi[0][i].item()

            print("(%.2f) %s" % (value, all_categories[category_index]))

            predictions.append([value, all_categories[category_index]])

# 测试

for evaluate_fn in [evaluateRNN, evaluateLSTM, evaluateGRU]:

    predict('Dovesky', evaluate_fn)

    predict('Jackson', evaluate_fn)

    predict('Satoshi', evaluate_fn)

完整代码版（方便复制来测试）

import torch

from io import open

import glob

import os

import unicodedata

import string

import random

import time

import math

import torch.nn as nn

import matplotlib.pyplot as plt

data_path = './data/names/'

all_letters = string.ascii_letters + " .,;'"

def unicodeToAscii(text):

    """

    Converts a Unicode string to an ASCII string.

    Args:

        text (str): The Unicode string to convert.

    Returns:

        str: The ASCII string.

    """

    return ''.join([

        unicodedata.normalize('NFKD', char)

        for char in text

        if not unicodedata.combining(char)

    ]).encode('ascii', 'ignore').decode('ascii')

def readLines(filename):

    lines = open(filename, encoding='utf-8').read().strip().split('\n')

    return [unicodeToAscii(line) for line in lines]

# 构建一个人名类别与具体人名对应关系的字典

category_lines = {}

# 构建所有类别的列表

all_categories = []

# 遍历所有的文件，使用glob.glob中可以利用正则表达式遍历

for filename in glob.glob(data_path + '*.txt'):

    category = os.path.splitext(os.path.basename(filename))[0]

    all_categories.append(category)

    lines = readLines(filename)

    # 将类别与人名对应关系存储到字典中

    category_lines[category] = lines

n_categories = len(all_categories)

def lineToTensor(line):

    # 首先初始化一个全0的张量，大小为len(line) * 1 * n_letters

    # 代表人名每个字母都用一个(1 * n_letters)的one-hot向量表示

    tensor = torch.zeros(len(line), 1, len(all_letters))

    # 遍历人名的每个字母, 并搜索其在所有字母中的位置，将其对应的位置置为1

    for li, letter in enumerate(line):

        tensor[li][0][all_letters.find(letter)] = 1

    return tensor

class RNN(nn.Module):

    def __init__(self, input_size, hidden_size, output_size, num_layers=1):

        super(RNN, self).__init__()

        # input_size: 输入数据的特征维度

        # hidden_size: RNN隐藏层的最后一个维度

        # output_size: RNN网络最后线性层的输出维度

        # num_layers: RNN网络的层数

        self.input_size = input_size

        self.hidden_size = hidden_size

        self.output_size = output_size

        self.num_layers = num_layers

        self.rnn = nn.RNN(input_size, hidden_size, num_layers)

        self.linear = nn.Linear(hidden_size, output_size)

        self.softmax = nn.LogSoftmax(dim=-1)

    def forward(self, input1, hidden):

        # input: 人名分类器中的输入张量，形状是1*n_letters

        # hidden: 代表隐藏层张量，形状是self.num_layers*1*hidden_size

        # 输入到RNN中的张量要求是三维张量，所以需要用unsqueeze()函数扩充维度

        input1 = input1.unsqueeze(0)

        rr, hn = self.rnn(input1, hidden)

        # 将RNN中获得的结果通过线性层变换和softmax函数输出

        return self.softmax(self.linear(rr)), hn

    def initHidden(self):

        # 初始化隐藏层张量，全0张量，维度是3

        return torch.zeros(self.num_layers, 1, self.hidden_size)

class LSTM(nn.Module):

    def __init__(self, input_size, hidden_size, output_size, num_layers=1):

        super(LSTM, self).__init__()

        self.input_size = input_size

        self.hidden_size = hidden_size

        self.output_size = output_size

        self.num_layers = num_layers

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

        self.linear = nn.Linear(hidden_size, output_size)

        self.softmax = nn.LogSoftmax(dim=-1)

    def forward(self, input1, hidden, c):

        # 注意：LSTM网络的输入有三个张量，不能忘记细胞状态C

        input1 = input1.unsqueeze(0)

        rr, (hn, cn) = self.lstm(input1, (hidden, c))

        return self.softmax(self.linear(rr)), hn, cn

    def initHiddenAndC(self):

        c = hidden = torch.zeros(self.num_layers, 1, self.hidden_size)

        return hidden, c

class GRU(nn.Module):

    def __init__(self, input_size, hidden_size, output_size, num_layers=1):

        super(GRU, self).__init__()

        self.input_size = input_size

        self.hidden_size = hidden_size

        self.output_size = output_size

        self.num_layers = num_layers

        self.gru = nn.GRU(input_size, hidden_size, num_layers)

        self.linear = nn.Linear(hidden_size, output_size)

        self.softmax = nn.LogSoftmax(dim=-1)

    def forward(self, input1, hidden):

        input1 = input1.unsqueeze(0)

        rr, hn = self.gru(input1, hidden)

        return self.softmax(self.linear(rr)), hn

    def initHidden(self):

        return torch.zeros(self.num_layers, 1, self.hidden_size)

# 参数

input_size = len(all_letters)

n_hidden = 128

output_size = n_categories

hidden = c = torch.zeros(1, 1, n_hidden)

rnn = RNN(input_size, n_hidden, output_size)

lstm = LSTM(input_size, n_hidden, output_size)

gru = GRU(input_size, n_hidden, output_size)

def categoryFromOutput(output):

    # 从输出中获取最大值和最大值的索引

    top_n, top_i = output.topk(1)

    category_i = top_i[0].item()

    return all_categories[category_i], category_i

# category, category_i = categoryFromOutput(rnn_output)

# print("category:", category)

# print("category_i:", category_i)

def randomTrainingExample():

    # 随机选择一个类别

    category = random.choice(all_categories)

    # 从该类别中随机选择一个人名

    line = random.choice(category_lines[category])

    # 将人名转换为张量

    category_tensor = torch.tensor([all_categories.index(category)], dtype=torch.long)

    line_tensor = lineToTensor(line)

    return category, line, category_tensor, line_tensor

# 构建传统RNN训练函数

criterion = nn.NLLLoss()

learning_rate = 0.005

def trainRNN(category_tensor, line_tensor):

    hidden = rnn.initHidden()

    rnn.zero_grad()

    output = None

    for i in range(line_tensor.size()[0]):

        output, hidden = rnn(line_tensor[i], hidden)

    # rnn对象由nn.RNN实例化得到，最终输出得到的是三维张量，为了满足category_tensor的维度要求，需要将其转换为二维张量

    loss = criterion(output.squeeze(0), category_tensor)

    loss.backward()

    # 更新参数

    for p in rnn.parameters():

        p.data.add_(-learning_rate, p.grad.data)

    return output, loss.item()

def trainLSTM(category_tensor, line_tensor):

    hidden, c = lstm.initHiddenAndC()

    lstm.zero_grad()

    output = None

    for i in range(line_tensor.size()[0]):

        output, hidden, c = lstm(line_tensor[i], hidden, c)

    loss = criterion(output.squeeze(0), category_tensor)

    loss.backward()

    for p in lstm.parameters():

        p.data.add_(-learning_rate, p.grad.data)

    return output, loss.item()

def trainGRU(category_tensor, line_tensor):

    hidden = gru.initHidden()

    gru.zero_grad()

    output = None

    for i in range(line_tensor.size()[0]):

        output, hidden = gru(line_tensor[i], hidden)

    loss = criterion(output.squeeze(0), category_tensor)

    loss.backward()

    for p in gru.parameters():

        p.data.add_(-learning_rate, p.grad.data)

    return output, loss.item()

# 构建时间计算函数

def timeSince(since):

    now = time.time()

    s = now - since

    m = math.floor(s / 60)

    s -= m * 60

    return "%dm %ds" % (m, s)

# 完整的训练函数

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

n_iters = 1000

print_every = 50

plot_every = 10

def train(train_type_fn):

    # 每个制图间隔损失保存列表

    all_losses = []

    # 获得开始的时间戳

    start = time.time()

    current_loss = 0

    for iter in range(1, n_iters + 1):

        category, line, category_tensor, line_tensor = randomTrainingExample()

        output, loss = train_type_fn(category_tensor, line_tensor)

        current_loss += loss

        if iter % print_every == 0:

            guess, guess_i = categoryFromOutput(output)

            correct = "✓" if guess == category else "✗ (%s)" % category

            print("%d %d%% (%s) %.4f %s / %s %s" % (iter, iter / n_iters * 100, timeSince(start), loss, line, guess, correct))

        if iter % plot_every == 0:

            all_losses.append(current_loss / plot_every)

            current_loss = 0

    return all_losses, int(time.time() - start)

# 训练并制作对比图

all_losses1, period1 = train(trainRNN)

all_losses2, period2 = train(trainLSTM)

all_losses3, period3 = train(trainGRU)

plt.figure(0)

plt.plot(all_losses1, label='RNN')

plt.plot(all_losses2, color='red', label='LSTM')

plt.plot(all_losses3, color='green', label='GRU')

plt.legend(loc='upper left')

plt.figure(1)

x_data = ['RNN', 'LSTM', 'GRU']

y_data = [period1, period2, period3]

plt.bar(range(len(x_data)), y_data, color='green', tick_label=x_data)

# 构建评估函数

def evaluateRNN(line_tensor):

    output = None

    hidden = rnn.initHidden()

    for i in range(line_tensor.size()[0]):

        output, hidden = rnn(line_tensor[i], hidden)

    return output.squeeze(0)

def evaluateLSTM(line_tensor):

    output = None

    hidden, c = lstm.initHiddenAndC()

    for i in range(line_tensor.size()[0]):

        output, hidden, c = lstm(line_tensor[i], hidden, c)

    return output.squeeze(0)

def evaluateGRU(line_tensor):

    output = None

    hidden = gru.initHidden()

    for i in range(line_tensor.size()[0]):

        output, hidden = gru(line_tensor[i], hidden)

    return output.squeeze(0)

# 构建预测函数

def predict(input_line, evaluate, n_predictions=3):

    print("\n> %s" % input_line)

    with torch.no_grad():

        output = evaluate(lineToTensor(input_line))

        topv, topi = output.topk(n_predictions, 1, True)

        predictions = []

        for i in range(n_predictions):

            value = topv[0][i].item()

            category_index = topi[0][i].item()

            print("(%.2f) %s" % (value, all_categories[category_index]))

            predictions.append([value, all_categories[category_index]])

# 调用试试

for evaluate_fn in [evaluateRNN, evaluateLSTM, evaluateGRU]:

    predict('Dovesky', evaluate_fn)

    predict('Jackson', evaluate_fn)

    predict('Satoshi', evaluate_fn)