加载中...

GRU简介

一、GRU介绍

  GRU是LSTM网络的一种效果很好的变体,它较LSTM网络的结构更加简单,而且效果也很好,因此也是当前非常流形的一种网络。GRU既然是LSTM的变体,因此也是可以解决RNN网络中的长依赖问题。

  GRU的参数较少,因此训练速度更快,GRU能够降低过拟合的风险。

  在LSTM中引入了三个门函数:输入门、遗忘门和输出门来控制输入值、记忆值和输出值。而在GRU模型中只有两个门:分别是更新门和重置门。具体结构如下图所示:


·
图中的zt和rt分别表示更新门和重置门。更新门用于控制前一时刻的状态信息被带入到当前状态中的程度,更新门的值越大说明前一时刻的状态信息带入越多。重置门控制前一状态有多少信息被写入到当前的候选集 h~t


二、GRU与LSTM的比较

  1. GRU相比于LSTM少了输出门,其参数比LSTM少。
  2. GRU在复调音乐建模和语音信号建模等特定任务上的性能和LSTM差不多,在某些较小的数据集上,GRU相比于LSTM表现出更好的性能。
  3. LSTM比GRU严格来说更强,因为它可以很容易地进行无限计数,而GRU却不能。这就是GRU不能学习简单语言的原因,而这些语言是LSTM可以学习的。
  4. GRU网络在首次大规模的神经网络机器翻译的结构变化分析中,性能始终不如LSTM。

三、GRU的API

rnn = nn.GRU(input_size, hidden_size, num_layers, bias, batch_first, dropout, bidirectional)

初始化:
input_size: input的特征维度
hidden_size: 隐藏层的宽度
num_layers: 单元的数量(层数),默认为1,如果为2以为着将两个GRU堆叠在一起,当成一个GRU单元使用。
bias: True or False,是否使用bias项,默认使用
batch_first: Ture or False, 默认的输入是三个维度的,即:(seq, batch, feature),第一个维度是时间序列,第二个维度是batch,第三个维度是特征。如果设置为True,则(batch, seq, feature)。即batch,时间序列,每个时间点特征。
dropout:设置隐藏层是否启用dropout,默认为0
bidirectional:True or False, 默认为False,是否使用双向的GRU,如果使用双向的GRU,则自动将序列正序和反序各输入一次。
输入:
rnn(input, h_0)

输出:
output, hn = rnn(input, h0)

形状的和LSTM差不多,也有双向

四、情感分类demo修改成GRU

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch import optim
import os
import re
import pickle
import numpy as np
from torch.utils.data import Dataset, DataLoader
from tqdm import tqdm


dataset_path = r'C:\Users\ci21615\Downloads\aclImdb_v1\aclImdb'
MAX_LEN = 500

def tokenize(text):
    """
    分词,处理原始文本
    :param text:
    :return:
    """
    fileters = ['!', '"', '#', '$', '%', '&', '\(', '\)', '\*', '\+', ',', '-', '\.', '/', ':', ';', '<', '=', '>', '\?', '@'
        , '\[', '\\', '\]', '^', '_', '`', '\{', '\|', '\}', '~', '\t', '\n', '\x97', '\x96', '”', '“', ]
    text = re.sub("<.*?>", " ", text, flags=re.S)
    text = re.sub("|".join(fileters), " ", text, flags=re.S)
    return [i.strip() for i in text.split()]
  
 class ImdbDataset(Dataset):
    """
    准备数据集
    """
    def __init__(self, mode):
        super(ImdbDataset, self).__init__()
        if mode == 'train':
            text_path = [os.path.join(dataset_path, i) for i in ['train/neg', 'train/pos']]
        else:
            text_path = [os.path.join(dataset_path, i) for i in ['test/neg', 'test/pos']]
        self.total_file_path_list = []
        for i in text_path:
            self.total_file_path_list.extend([os.path.join(i, j) for j in os.listdir(i)])
     def __getitem__(self, item):
        cur_path = self.total_file_path_list[item]
        cur_filename = os.path.basename(cur_path)
        # 获取标签
        label_temp = int(cur_filename.split('_')[-1].split('.')[0]) - 1
        label = 0 if label_temp < 4 else 1
        text = tokenize(open(cur_path, encoding='utf-8').read().strip())
        return label, text
     def __len__(self):
        return len(self.total_file_path_list)
 
  
 class Word2Sequence():
    UNK_TAG = 'UNK'
    PAD_TAG = 'PAD'
    UNK = 0
    PAD = 1
 
    def __init__(self):
        self.dict = {
            self.UNK_TAG: self.UNK,
            self.PAD_TAG: self.PAD
         }
        self.count = {} # 统计词频

    def fit(self, sentence):
        """
        把单个句子保存到dict中
        :return:
        """
        for word in sentence:
            self.count[word] = self.count.get(word, 0) + 1
 
    def build_vocab(self, min=5, max=None, max_feature=None):
        """
        生成词典
        :param min: 最小出现的次数
        :param max: 最大次数
        :param max_feature: 一共保留多少个词语
        :return:
        """
        # 删除词频小于min的word
        if min is not None:
            self.count = {word:value for word,value in self.count.items() if value > min}
        # 删除词频大于max的word
        if max is not None:
            self.count = {word:value for word,value in self.count.items() if value < max}
        # 限制保留的词语数
        if max_feature is not None:
            temp = sorted(self.count.items(), key=lambda x:x[-1],reverse=True)[:max_feature]
            self.count = dict(temp)
        for word in self.count:
            self.dict[word] = len(self.dict)
        # 得到一个反转的字典
        self.inverse_dict = dict(zip(self.dict.values(), self.dict.keys()))

    def transform(self, sentence, max_len=None):
        """
        把句子转化为序列
        :param sentence: [word1, word2...]
        :param max_len: 对句子进行填充或裁剪
        :return:
        """
        if max_len is not None:
            if max_len > len(sentence):
                sentence = sentence + [self.PAD_TAG] * (max_len - len(sentence)) # 填充
            if max_len < len(sentence):
                sentence = sentence[:max_len] # 裁剪
         return [self.dict.get(word, self.UNK) for word in sentence]
 
    def inverse_transform(self, indices):
       """
        把序列转化为句子
        :param indices: [1,2,3,4...]
        :return:
        """
        return [self.inverse_dict.get(idx) for idx in indices]
 
    def __len__(self):
        return len(self.dict)
 
 
def fit_save_word_sequence():
     """
     从数据集构建字典
     :return:
     """
     ws = Word2Sequence()
     train_path = [os.path.join(dataset_path, i) for i in ['train/neg', 'train/pos']]
     total_file_path_list = []
     for i in train_path:
         total_file_path_list.extend([os.path.join(i, j) for j in os.listdir(i)])
     for cur_path in tqdm(total_file_path_list, desc='fitting'):
         sentence = open(cur_path, encoding='utf-8').read().strip()
         res = tokenize(sentence)
         ws.fit(res)
     # 对wordSequesnce进行保存
     ws.build_vocab(min=10)
     # pickle.dump(ws, open('./lstm_model/ws.pkl', 'wb'))
     return ws
 
def get_dataloader(mode='train', batch_size=20, ws=None):
     """
     获取数据集,转换成词向量后的数据集
     :param mode:
     :return:
     """
     # 导入词典
     # ws = pickle.load(open('./model/ws.pkl', 'rb'))
     # 自定义collate_fn函数
     def collate_fn(batch):
         """
         batch是list,其中是一个一个元组,每个元组是dataset中__getitem__的结果
         :param batch:
         :return:
         """
         batch = list(zip(*batch))
         labels = torch.LongTensor(batch[0])
         texts = batch[1]
         # 获取每个文本的长度
         lengths = [len(i) if len(i) < MAX_LEN else MAX_LEN for i in texts]
         # 每一段文本句子都转换成了n个单词对应的数字组成的向量,即500个单词数字组成的向量
         temp = [ws.transform(i, MAX_LEN) for i in texts]
         texts = torch.LongTensor(temp)
         del batch
         return labels, texts, lengths
     dataset = ImdbDataset(mode)
     dataloader = DataLoader(dataset=dataset, batch_size=batch_size, shuffle=True, collate_fn=collate_fn)
     return dataloader
 
 
 class ImdbLstmModel(nn.Module):
 
     def __init__(self, ws):
         super(ImdbLstmModel, self).__init__()
         self.hidden_size = 64   # 隐藏层神经元的数量,即每一层有多少个LSTM单元
         self.embedding_dim = 200    # 每个词语使用多长的向量表示
         self.num_layer = 1  # 即RNN的中LSTM单元的层数
         self.bidriectional = True  # 是否使用双向LSTM,默认是False,表示双向LSTM,也就是序列从左往右算一次,从右往左又算一次,这样就可以两倍的输出
         self.num_directions = 2 if self.bidriectional else 1 # 是否双向取值,双向取值为2,单向取值为1
         self.dropout = 0.5  # dropout的比例,默认值为0。dropout是一种训练过程中让部分参数随机失活的一种方式,能够提高训练速度,同时能够解决过拟合的问题。这里是在LSTM的最后一层,对每个输出进行dropout
         # 每个句子长度为500
         # ws = pickle.load(open('./model/ws.pkl', 'rb'))
         print(len(ws))
         self.embedding = nn.Embedding(len(ws), self.embedding_dim)
         # self.lstm = nn.LSTM(self.embedding_dim,self.hidden_size,self.num_layer,bidirectional=self.bidriectional,dropout=self.dropout)
         self.gru = nn.GRU(input_size=self.embedding_dim, hidden_size=self.hidden_size, bidirectional=self.bidriectional)
 
         self.fc = nn.Linear(self.hidden_size * self.num_directions, 20)
         self.fc2 = nn.Linear(20, 2)
 
     def init_hidden_state(self, batch_size):
         """
         初始化 前一次的h_0(前一次的隐藏状态)和c_0(前一次memory)
         :param batch_size:
         :return:
         """
         h_0 = torch.rand(self.num_layer * self.num_directions, batch_size, self.hidden_size)
         return h_0
 
     def forward(self, input):
         # 句子转换成词向量
         x = self.embedding(input)
         # 如果batch_first为False的话转换一下seq_len和batch_size的位置
         x = x.permute(1,0,2)    # [seq_len, batch_size, embedding_num]
         # 初始化前一次的h_0(前一次的隐藏状态)和c_0(前一次memory)
         h_0 = self.init_hidden_state(x.size(1))    # [num_layers * num_directions, batch, hidden_size]
         output, h_n = self.gru(x, h_0)

         # 只要最后一个lstm单元处理的结果,这里多去的hidden state
         out = torch.cat([h_n[-2, :, :], h_n[-1, :, :]], dim=-1)
         out = self.fc(out)
         out = F.relu(out)
         out = self.fc2(out)
         return F.log_softmax(out, dim=-1)

 
 train_batch_size = 64
 test_batch_size = 5000
 
 def train(epoch, ws):
     """
     训练
     :param epoch: 轮次
     :param ws: 字典
     :return:
     """
     mode = 'train'
     imdb_lstm_model = ImdbLstmModel(ws)
     optimizer = optim.Adam(imdb_lstm_model.parameters())
     for i in range(epoch):
         train_dataloader = get_dataloader(mode=mode, batch_size=train_batch_size, ws=ws)
         for idx, (target, input, input_length) in enumerate(train_dataloader):
             optimizer.zero_grad()
             output = imdb_lstm_model(input)
             loss = F.nll_loss(output, target)
             loss.backward()
             optimizer.step()
 
             pred = torch.max(output, dim=-1, keepdim=False)[-1]
             acc = pred.eq(target.data).numpy().mean() * 100.
             print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}\t ACC: {:.6f}'.format(i, idx * len(input), len(train_dataloader.dataset),
                                                                                     100. * idx / len(train_dataloader), loss.item(), acc))
     torch.save(imdb_lstm_model.state_dict(), 'model/gru_model.pkl')
     torch.save(optimizer.state_dict(), 'model/gru_optimizer.pkl')
 
 
def test(ws):
     mode = 'test'
     # 载入模型
     lstm_model = ImdbLstmModel(ws)
     lstm_model.load_state_dict(torch.load('model/lstm_model.pkl'))
     optimizer = optim.Adam(lstm_model.parameters())
     optimizer.load_state_dict(torch.load('model/lstm_optimizer.pkl'))
     lstm_model.eval()
     test_dataloader = get_dataloader(mode=mode, batch_size=test_batch_size, ws=ws)
     with torch.no_grad():
         for idx, (target, input, input_length) in enumerate(test_dataloader):
             output = lstm_model(input)
             test_loss = F.nll_loss(output, target, reduction='mean')
             pred = torch.max(output, dim=-1, keepdim=False)[-1]
             correct = pred.eq(target.data).sum()
             acc = 100. * pred.eq(target.data).cpu().numpy().mean()
             print('idx: {} Test set: Avg. loss: {:.4f}, Accuracy: {}/{} ({:.2f}%)\n'.format(idx, test_loss, correct, target.size(0), acc))
 
 
if __name__ == '__main__':
     # 构建字典
     ws = fit_save_word_sequence()
     # 训练
     train(10, ws)
     # 测试
     # test(ws)

结果展示:

posted @ 2023-03-19 10:45  星芒易逝  阅读(1303)  评论(0编辑  收藏  举报