LSTM进行情感分析
LSTM进行情感分析的复现–pytorch的实现
关于TextCNN的复现参考本文章
TextCNN的复现–pytorch实现 - 知乎 (zhihu.com)
接下来主要是对代码内容的详解，完整代码将在文章末尾给出。
使用的数据集为电影评论数据集，其中正面数据集5000条左右，负面的数据集也为5000条。
pyroch的基本训练过程：
加载训练集–构建模型–模型训练–模型评价
首先，是要对数据集进行加载，在对数据集加载时候需要继承一下Dataset类，代码如下
import re
from collections import Counter
from collections import OrderedDict
import gensim
import torch.nn
from sympy.parsing.sympy_parser import _flatten
from torch.utils.data import Dataset
import numpy as np
import pandas as pd
from torchtext.vocab import vocab
class Data_loader(Dataset):
def __init__(self, file_pos, file_neg, model_path, word2_vec=False):
self.file_pos = file_pos
self.file_neg = file_neg
if word2_vec:
self.x_train, self.y_train = self.get_word2vec(model_path)
else:
self.x_train, self.y_train, self.dictionary = self.pre_process()
def __getitem__(self, idx):
data = self.x_train[idx]
label = self.y_train[idx]
data = torch.tensor(data)
label = torch.tensor(label)
return data, label
def __len__(self):
return len(self.x_train)
def clean_sentences(self, string):
string = re.sub(r"[^A-Za-z0-9(),!?\'\`]", " ", string)
string = re.sub(r"\'s", " \'s", string)
string = re.sub(r"\'ve", " \'ve", string)
string = re.sub(r"n\'t", " n\'t", string)
string = re.sub(r"\'re", " \'re", string)
string = re.sub(r"\'d", " \'d", string)
string = re.sub(r"\'ll", " \'ll", string)
string = re.sub(r",", " , ", string)
string = re.sub(r"!", " ! ", string)
string = re.sub(r"\(", " \( ", string)
string = re.sub(r"\)", " \) ", string)
string = re.sub(r"\?", " \? ", string)
string = re.sub(r"\s{2,}", " ", string)
return string.strip().lower()
def load_data_and_labels(self):
positive_examples = list(open(self.file_pos, "r", encoding="utf-8").readlines())
positive_examples = [s.strip() for s in positive_examples] # 对评论数据删除每一行数据的\t，\n
negative_examples = list(open(self.file_neg, "r", encoding="utf-8").readlines())
negative_examples = [s.strip() for s in negative_examples] # 对评论数据删除每一行数据的\t，\n
x_text = positive_examples + negative_examples
x_text = [self.clean_sentences(_) for _ in x_text]
positive_labels = [1 for _ in positive_examples] # 正样本数据为1
negative_labels = [0 for _ in negative_examples] # 负样本数据为0
y = np.concatenate((positive_labels, negative_labels), axis=0)
return x_text, y.T # 返回的是dataframe对象，[0]data[0]为文本数据，data[1]为标签
def pre_process(self):
'''
加载数据，并对之前使用的数据进行打乱返回,同时根据训练集和测试集的比列进行划分,默认百分80和百分20
:return:测试数据、训练数据、以及生成的词汇表
'''
x_data, y_label = self.load_data_and_labels()
max_document_length = max(len(x.split(' ')) for x in x_data)
voc = []
word_split = []
[voc.extend(x.split()) for x in x_data] # 生成词典
[word_split.append(x.split()) for x in x_data]
if len(voc) != 0:
ordere_dict = OrderedDict(sorted(Counter(_flatten(voc)).items(), key=lambda x: x[1], reverse=True))
# 把文档映射成词汇的索引序列
dictionary = vocab(ordere_dict)
x_data = []
for words in word_split:
x = list(dictionary.lookup_indices(words))
temp_pos = max_document_length - len(x)
if temp_pos != 0:
for i in range(1, temp_pos + 1):
x.extend([0])
x_data.append(x)
x_data = np.array(x_data)
np.random.seed(10)
# 将标签打乱顺序，返回索引
shuffle_indices = np.random.permutation(np.arange(len(y_label)))
x_shuffled = x_data[shuffle_indices]
y_shuffled = y_label[shuffle_indices]
return x_shuffled, y_shuffled, dictionary
def get_word2vec(self, model_path):
model = gensim.models.Word2Vec.load(model_path)
x_data, y_label = self.load_data_and_labels()
word_split = []
[word_split.append(x.split()) for x in x_data]
sentence_vectors = []
for sentence in word_split:
sentence_vector = []
for word in sentence:
try:
v = model.wv.get_index(word)
except Exception as e:
v = np.random.randint(0, 71289)
sentence_vector.append(int(v))
sentence_vectors.append(sentence_vector) # 获取到句子单词在词表中的位置
max_document_length = max(len(x) for x in sentence_vectors)
for vector in sentence_vectors:
for i in range(1, max_document_length - len(vector) + 1):
vector.append(int(71289))
vector_data = np.asarray(sentence_vectors)
# vector_data = sentence_vectors
np.random.seed(10)
# 将标签打乱顺序，返回索引
shuffle_indices = np.random.permutation(np.arange(len(y_label)))
x_shuffled = vector_data[shuffle_indices]
y_shuffled = y_label[shuffle_indices]
return x_shuffled, y_shuffled
# return vector_data, y_label
上述代码中的__init__ 、getitem 、len是必须要继承实现的方法，clean_sentence是对读取的数据进行清洗，load_data_and_label是加载数据且返回清洗过后的数据以及数据标签。pre_process是对数据进行编码，原始的数据是英文数据，因此需要对其进行分词、编码，最后返回的数据将是数字，一行数据就是一句评论。
例如：
I like this movie
在对其进行编码返回后将是 0 1 2 3,0对应的为I，1对应的为like以此类推。
在这儿设置的每个词的维度是256维。
接下来就是LSTM模型的构建
class LSTM_RNN(nn.Module):
def __init__(self, num_embeddings=-1, drop_rate=0.8, embedding_dim=256, hidden_size=64, output_size=1, num_layers=2):
super().__init__()
self.num_layers = num_layers
self.hidden_size = hidden_size
self.num_embeddings = num_embeddings # num_embeddings为单词的维度
self.embedding = nn.Embedding(num_embeddings, embedding_dim)
self.lstm = nn.LSTM(embedding_dim, hidden_size, num_layers, bidirectional=True, batch_first=True)
self.compute = nn.Linear(hidden_size, output_size)
self.drop = nn.Dropout(drop_rate)
self.sigmod = nn.Sigmoid()
def forward(self, x, hidden):
"""
x: 本次的输入，其size为(batch_size, 200)，200为句子长度
hidden: 上一时刻的Hidden State和Cell State。类型为tuple: (h, c),
其中h和c的size都为(n_layers, batch_size, hidden_dim), 即(2, 200, 512)
"""
if self.num_embeddings > 0:
x = self.embedding(x)
batch_size = x.size(0)
x, hidden = self.lstm(x, hidden) # _x is input, size (seq_len, batch, input_size)
# s, b, h = x.shape # x is output, size (seq_len, batch, hidden_size)
x = x.contiguous().view(-1, self.hidden_size)
x = self.drop(x)
x = self.compute(x)
# x = x.view(s, b, -1)
predict = self.sigmod(x)
predict = predict.view(batch_size, -1)
out = predict[:, -1] # 取得是最后一个单词的概率
return out, hidden
def init_hidden(self, batch_size):
"""
初始化隐状态：第一次送给LSTM时，没有隐状态，所以要初始化一个
这里的初始化策略是全部赋0。
这里之所以是tuple，是因为LSTM需要接受两个隐状态hidden state和cell state
"""
hidden = (torch.zeros(self.num_layers*2, batch_size, self.hidden_size).to('cpu'),
torch.zeros(self.num_layers*2, batch_size, self.hidden_size).to('cpu')
)
return hidden
在构建模型时候需要继承nn.moudule,同时要实现__init__、以及forward方法，可以看作init在定义各个层，forward在对各个层之间来进行连接。
接下来就是对模型进行训练，代码如下所示:
import torch
from torch import nn
from torch.utils.data import DataLoader
import torch.nn
from LSTM import LSTM_RNN
from TextCNN import TextCNN
from dataLoader import Data_loader
batch_size = 830
# num_classes = 2
file_pos = 'E:\\PostGraduate\\Paper_review\\pytorch_TextCnn/data/rt-polarity.pos'
file_neg = 'E:\\PostGraduate\\Paper_review\\pytorch_TextCnn/data/rt-polarity.neg'
word2vec_path = 'E:\\PostGraduate\\Paper_review\\pytorch_TextCnn/word2vec1.model'
train_data = Data_loader(file_pos, file_neg, word2vec_path)
train_size = int(len(train_data) * 0.8)
test_size = len(train_data) - train_size
train_dataset, test_dataset = torch.utils.data.random_split(train_data, [train_size, test_size])
train_iter = DataLoader(train_dataset, batch_size=batch_size, shuffle=True)
test_iter = DataLoader(test_dataset, batch_size=batch_size, shuffle=True)
# model = TextCNN(num_classes, embeddings_pretrained=True)
model = LSTM_RNN(num_embeddings=18764)
# 开始训练
epoch = 100 # 训练轮次
optmizer = torch.optim.Adam(model.parameters(), lr=0.001)
# optmizer = torch.optim.SGD(model.parameters(),lr=0.01,momentum=0.04)
train_losses = []
train_counter = []
test_losses = []
log_interval = 5
test_counter = [i * len(train_iter.dataset) for i in range(epoch + 1)]
device = 'cpu'
def train_loop(n_epochs, optimizer, model, train_loader, device, test_iter):
for epoch in range(1, n_epochs + 1):
print("开始第{}轮训练".format(epoch))
model.train()
correct = 0
for i, data in enumerate(train_loader):
# print(i)
optimizer.zero_grad()
(text_data, label) = data
text_data = text_data.to(device)
label = label.to(device)
label = label.long()
# print(len(text_data))
h = model.init_hidden(len(text_data)) # 初始化第一个Hidden_state
output, h = model(text_data, h)
# print(torch.mean(output))
loss_func = nn.BCELoss()
# output = output.long()
loss = loss_func(output, label.float())
loss.backward()
optimizer.step()
pred = [1 if x >= 0.5 else 0 for x in output] # 返回的是列表
for index in range(0, len(pred)):
if pred[index] == label[index]:
correct += 1
if i % log_interval == 0:
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
epoch, i * len(text_data), len(train_loader.dataset),
100. * i / len(train_loader), loss.item()))
train_losses.append(loss.item())
train_counter.append(
(i * 64) + ((epoch - 1) * len(train_loader.dataset)))
torch.save(model.state_dict(), './model.pth')
torch.save(optimizer.state_dict(), './optimizer.pth')
# if 100. * correct / len(train_loader.dataset)<94:
print("Accuracy: {}/{} ({:.0f}%)\n".format(correct, len(train_loader.dataset),
100. * correct / len(train_loader.dataset)))
test_loop(model, device, test_iter)
# else:
# break
# model.eval()
# PATH = 'E:\\PostGraduate\\Paper_review\\pytorch_TextCnn\\LSTM/model.pth'
# dictionary = torch.load(PATH)
# model.load_state_dict(dictionary)
def test_loop(model, device, test_iter):
model.eval()
test_loss = 0
correct = 0
with torch.no_grad():
for data, target in test_iter:
data = data.to(device)
target = target.to(device)
h = model.init_hidden(len(data)) # 初始化第一个Hidden_state
output, h = model(data, h)
loss_func = nn.BCEWithLogitsLoss()
loss = loss_func(output, target.float())
test_loss += loss
pred = [1 if x >= 0.5 else 0 for x in output] # 返回的是列表
for index in range(0, len(pred)):
if pred[index] == target[index]:
correct += 1
test_loss /= len(test_iter.dataset)
test_losses.append(test_loss)
print('\nTest set: Avg. loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
test_loss, correct, len(test_iter.dataset),
100. * correct / len(test_iter.dataset)))
train_loop(epoch, optmizer, model, train_iter, device, test_iter)
PATH = 'E:\\PostGraduate\\Paper_review\\pytorch_TextCnn\\LSTM/model.pth'
model = LSTM_RNN(num_embeddings=18764)
dictionary = torch.load(PATH)
model.load_state_dict(dictionary)
test_loop(model, device, test_iter)
在上述中，首先会对数据集加载进来，然后分为80%的训练集和20%的测试集，定义使用的优化器为adam。同时在训练的过程中会对优化器、损失函数等信息进行保存。
训练结果如下所示：
完整代码链接，后续会使用其他数据来对模型进行测试。
/model.pth’
model = LSTM_RNN(num_embeddings=18764)
dictionary = torch.load(PATH)
model.load_state_dict(dictionary)
test_loop(model, device, test_iter)
在上述中，首先会对数据集加载进来，然后分为80%的训练集和20%的测试集，定义使用的优化器为adam。同时在训练的过程中会对优化器、损失函数等信息进行保存。
训练结果如下所示：
[外链图片转存中...(img-KOExQS59-1706516526357)]
完整代码链接，后续会使用其他数据来对模型进行测试。
[木南/TextCNN (gitee.com)](https://gitee.com/nanwang-crea/text-cnn)