Pytorch官方教程：用RNN实现字符级的生成任务

数据处理

传送门：官方教程

数据从上面下载。本次的任务用到的数据和第一次一样，还是18个不同国家的不同名字。

但这次需要根据这些数据训练一个模型，给定国家和名字的首字母时，模型可以自动生成名字。

 

首先还是对数据进行预处理，和第一个任务一样，利用Unicode将不同国家的名字采用相同的编码方式，因为要生成名字，所以需要加上一个终止符，具体作用后面会提到。

import string
import os
import glob
import unicodedata


all_letters = string.ascii_letters + " .,;'-"  # string.ascii_letters的作用是生成所有的英文字母
n_letters = len(all_letters) + 1     # 多加的1是指EOS

n_hidden = 128
n_categories = 18

def find_files(path):
    """
    :param path:文件路径
    :return: 文件列表地址
    """
    return glob.glob(path)     # glob模块提供了一个函数用于从目录通配符搜索中生成文件列表:


def unicode_to_ascii(str):
    """
    :param str:名字
    :return:返回均采用NFD编码方式的名字
    """
    return ''.join(
      c for c in unicodedata.normalize('NFD', str)    # 对文字采用相同的编码方式
      if unicodedata.category(c) != 'Mn' and c in all_letters 
    )


def read_lines(files_list):
    """
    读取每个文件的内容
    :param files_list:文件所在地址列表
    :return:{国家：名字列表}
    """
    category_lines = {}
    all_categories = []
    for file in files_list:
        # os.path.splitext:分割路径，返回路径名和文件扩展名的元组
        # os.path.basename:返回文件名
        category = os.path.splitext(os.path.basename(file))[0]
        line = [unicode_to_ascii(line) for line in open(file)]
        all_categories.append(category)
        category_lines[category] = line
    return all_categories, category_lines

　　

 

如同官方教程上的这张图片，当我们训练的数据是Kasparov时，当我们输入K，我们希望它输出a；输入a，希望输出s；... ；输入v，希望输出终止符EOS。所有对于每一个input，都需要构建一个output作为比对，计算损失。

因为名字和国家有关，所以国家类别也要作为input的一部分进行训练，下面的代码实现的是将训练所需的数据转化为张量。

import torch


def category_to_tensor(category):
    """
    将类别转换成张量
    :param category:类别
    :return:张量
    """
    li = all_categories.index(category)
    tensor = torch.zeros(1, n_categories)
    tensor[0][li] = 1
    return tensor.to(device)


def input_to_tensor(input):
    """
    将输入进行one-hot编码
    :param word: 单词
    :return: 张量
    """
    tensor = torch.zeros(len(input), 1, n_letters)
    for i, letter in enumerate(input):
        tensor[i][0][all_letters.find(letter)] = 1
    return tensor.to(device)


def target_to_tensor(input):
    """
    对目标输出进行one-hot编码，即为从第二个字母开始至结束字母的索引，以及EOS的索引
    :param input:单词
    :return:张量
    """
    letter_indexes = [all_letters.find(input[i]) for i in range(1, len(input))]
    letter_indexes.append(n_letters - 1)    # 最后一位的索引是EOS
    return torch.LongTensor(letter_indexes).to(device)

　　

模型构建

 

class RNN(nn.Module):
    def __init__(self, input_size, hidden_size, output_size):
        super(RNN, self).__init__()

        self.hidden_size = hidden_size

        self.i2h = nn.Linear(n_categories + input_size + hidden_size, hidden_size)
        self.i2o = nn.Linear(n_categories + input_size + hidden_size, output_size)
        self.o2o = nn.Linear(hidden_size + output_size, output_size)
        self.dropout = nn.Dropout(0.1)
        self.softmax = nn.LogSoftmax(dim=1)  # dim=1表示对第1维度的数据进行logsoftmax操作

    def forward(self, category, input, hidden):
        input_tmp = torch.cat((category, input, hidden), 1)
        hidden = self.i2h(input_tmp)
        output = self.i2o(input_tmp)
        output_tmp = torch.cat((hidden, output), 1)
        output = self.o2o(output_tmp)
        output = self.dropout(output)
        output = self.softmax(output)
        return output, hidden

    def init_hidden(self):   # 隐藏层初始化0操作
        return torch.zeros(1, self.hidden_size).to(device)

　　

模型训练

 训练的过程依旧是迭代多次，每次从数据中随机选择，并将所需数据转换为张量。

import random


def random_choice(obj):
    return obj[random.randint(0, len(obj)-1)]


def random_training_example():
    category = random_choice(all_categories)
    input = random_choice(category_lines[category])
    category_tensor = category_to_tensor(category)
    input_tensor = input_to_tensor(input)
    target_tensor = target_to_tensor(input)
    return category_tensor, input_tensor, target_tensor

训练代码如下：

import torch.nn as nn
import time
import matplotlib.pyplot as plt
import matplotlib.ticker as ticker


def train():
    rnn = RNN(n_letters, n_hidden, n_letters).to(device)

    loss = nn.NLLLoss()
    total_loss = 0
    all_losses = []
    epoch_num = 100000
    lr = 0.0005
    epoch_start_time = time.time()

    for epoch in range(epoch_num):
        rnn.zero_grad()
        train_loss = 0
        hidden = rnn.init_hidden()
        category_tensor, input_tensor, target_tensor = random_training_example()
        target_tensor.unsqueeze_(-1)

        for i in range(input_tensor.size()[0]):
            output, hidden = rnn(category_tensor, input_tensor[i], hidden)
            train_loss += loss(output, target_tensor[i])   # 每次的loss都需要计算
        train_loss.backward()

        for p in rnn.parameters():
            p.data.add_(p.grad.data, alpha=-lr)

        total_loss += train_loss.item() / input_tensor.size()[0]    # 该名字的平均损失

        if epoch % 5000 == 0:
            print('[%05d/%03d%%] %2.2f sec(s) Loss: %.4f' %
                  (epoch, epoch / epoch_num * 100, time.time()-epoch_start_time, train_loss.item() / input_tensor.size()[0]))
            epoch_start_time = time.time()

        if (epoch+1) % 500 == 0:
            all_losses.append(total_loss / 500)
            total_loss = 0

    torch.save(rnn.state_dict(), 'rnn_params.pkl')   # 保存模型的数据

    plt.figure()
    plt.plot(all_losses)
    plt.show()

 

模型预测

def predict(category, start_letter):
    rnn = RNN(n_letters, n_hidden, n_letters).to(device)
    rnn.load_state_dict(torch.load('rnn_params.pkl'))       # 加载模型训练所得到的参数
    max_length = 20                                         # 名字的最大长度
    with torch.no_grad():
        category_tensor = category_to_tensor(category)
        input = input_to_tensor(start_letter)
        hidden = rnn.init_hidden()

        output_name = start_letter

        for i in range(max_length):
            output, hidden = rnn(category_tensor, input[0], hidden)
            top_v, top_i = output.topk(1)     # 选出最大的值，返回其value和index
            top_i = top_i.item()
            if top_i == n_letters - 1:        # n-letters-1是EOS
                break
            else:
                letter = all_letters[top_i]
                output_name += letter
            input = input_to_tensor(letter)   # 更新input，继续循环迭代
    return output_name

 

 

完整代码

import string
import os
import glob
import unicodedata
import torch
import torch.nn as nn
import random
import time
import matplotlib.pyplot as plt
import matplotlib.ticker as ticker


device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

all_letters = string.ascii_letters + " .,;'-"  # string.ascii_letters的作用是生成所有的英文字母
n_letters = len(all_letters) + 1     # 多加的1是指EOS

n_hidden = 128
n_categories = 18


class RNN(nn.Module):
    def __init__(self, input_size, hidden_size, output_size):
        super(RNN, self).__init__()

        self.hidden_size = hidden_size

        self.i2h = nn.Linear(n_categories + input_size + hidden_size, hidden_size)
        self.i2o = nn.Linear(n_categories + input_size + hidden_size, output_size)
        self.o2o = nn.Linear(hidden_size + output_size, output_size)
        self.dropout = nn.Dropout(0.1)
        self.softmax = nn.LogSoftmax(dim=1)  # dim=1表示对第1维度的数据进行logsoftmax操作

    def forward(self, category, input, hidden):
        input_tmp = torch.cat((category, input, hidden), 1)
        hidden = self.i2h(input_tmp)
        output = self.i2o(input_tmp)
        output_tmp = torch.cat((hidden, output), 1)
        output = self.o2o(output_tmp)
        output = self.dropout(output)
        output = self.softmax(output)
        return output, hidden

    def init_hidden(self):   # 隐藏层初始化0操作
        return torch.zeros(1, self.hidden_size).to(device)


def find_files(path):
    """
    :param path:文件路径
    :return: 文件列表地址
    """
    return glob.glob(path)     # glob模块提供了一个函数用于从目录通配符搜索中生成文件列表:


def unicode_to_ascii(str):
    """
    :param str:名字
    :return:返回均采用NFD编码方式的名字
    """
    return ''.join(
      c for c in unicodedata.normalize('NFD', str)    # 对文字采用相同的编码方式
      if unicodedata.category(c) != 'Mn' and c in all_letters 
    )


def read_lines(files_list):
    """
    读取每个文件的内容
    :param files_list:文件所在地址列表
    :return:{国家：名字列表}
    """
    category_lines = {}
    all_categories = []
    for file in files_list:
        # os.path.splitext:分割路径，返回路径名和文件扩展名的元组
        # os.path.basename:返回文件名
        category = os.path.splitext(os.path.basename(file))[0]
        line = [unicode_to_ascii(line) for line in open(file)]
        all_categories.append(category)
        category_lines[category] = line
    return all_categories, category_lines


def category_to_tensor(category):
    """
    将类别转换成张量
    :param category:类别
    :return:张量
    """
    li = all_categories.index(category)
    tensor = torch.zeros(1, n_categories)
    tensor[0][li] = 1
    return tensor.to(device)


def input_to_tensor(input):
    """
    将输入进行one-hot编码
    :param word: 单词
    :return: 张量
    """
    tensor = torch.zeros(len(input), 1, n_letters)
    for i, letter in enumerate(input):
        tensor[i][0][all_letters.find(letter)] = 1
    return tensor.to(device)


def target_to_tensor(input):
    """
    对目标输出进行one-hot编码，即为从第二个字母开始至结束字母的索引，以及EOS的索引
    :param input:单词
    :return:张量
    """
    letter_indexes = [all_letters.find(input[i]) for i in range(1, len(input))]
    letter_indexes.append(n_letters - 1)    # 最后一位的索引是EOS
    return torch.LongTensor(letter_indexes).to(device)


def random_choice(obj):
    return obj[random.randint(0, len(obj)-1)]


def random_training_example():
    category = random_choice(all_categories)
    input = random_choice(category_lines[category])
    category_tensor = category_to_tensor(category)
    input_tensor = input_to_tensor(input)
    target_tensor = target_to_tensor(input)
    return category_tensor, input_tensor, target_tensor


def train():
    rnn = RNN(n_letters, n_hidden, n_letters).to(device)

    loss = nn.NLLLoss()
    total_loss = 0
    all_losses = []
    epoch_num = 100000
    lr = 0.0005
    epoch_start_time = time.time()

    for epoch in range(epoch_num):
        rnn.zero_grad()
        train_loss = 0
        hidden = rnn.init_hidden()
        category_tensor, input_tensor, target_tensor = random_training_example()
        target_tensor.unsqueeze_(-1)

        for i in range(input_tensor.size()[0]):
            output, hidden = rnn(category_tensor, input_tensor[i], hidden)
            train_loss += loss(output, target_tensor[i])   # 每次的loss都需要计算
        train_loss.backward()

        for p in rnn.parameters():
            p.data.add_(p.grad.data, alpha=-lr)

        total_loss += train_loss.item() / input_tensor.size()[0]    # 该名字的平均损失

        if epoch % 5000 == 0:
            print('[%05d/%03d%%] %2.2f sec(s) Loss: %.4f' %
                  (epoch, epoch / epoch_num * 100, time.time()-epoch_start_time, train_loss.item() / input_tensor.size()[0]))
            epoch_start_time = time.time()

        if (epoch+1) % 500 == 0:
            all_losses.append(total_loss / 500)
            total_loss = 0

    torch.save(rnn.state_dict(), 'rnn_params.pkl')   # 保存模型的数据

    plt.figure()
    plt.plot(all_losses)
    plt.show()


def predict(category, start_letter):
    rnn = RNN(n_letters, n_hidden, n_letters).to(device)
    rnn.load_state_dict(torch.load('rnn_params.pkl'))       # 加载模型训练所得到的参数
    max_length = 20                                         # 名字的最大长度
    with torch.no_grad():
        category_tensor = category_to_tensor(category)
        input = input_to_tensor(start_letter)
        hidden = rnn.init_hidden()

        output_name = start_letter

        for i in range(max_length):
            output, hidden = rnn(category_tensor, input[0], hidden)
            top_v, top_i = output.topk(1)     # 选出最大的值，返回其value和index
            top_i = top_i.item()
            if top_i == n_letters - 1:        # n-letters-1是EOS
                break
            else:
                letter = all_letters[top_i]
                output_name += letter
            input = input_to_tensor(letter)   # 更新input，继续循环迭代
    return output_name


if __name__ == '__main__':
    files_list = find_files('./names/*.txt')
    all_categories, category_lines = read_lines(files_list)
    #train()
    print(predict('English', 'V'))