基于深度学习的文本情感分析（Keras VS Pytorch）

  实验任务：利用不同的深度学习框架对微博短文本进行情感分析，并将情感分为三类，分别是正、负、中。

  所用语言及相应的工具包：Python 3.6, Keras 2.2.4, Torch 1.0.1

  数据分布： {'pos': 712, 'neu': 768, 'neg': 521}

  技术路线：

本次实验利用词向量来表示文本，单条文本的形状为[50, 100].

 利用Keras对处理好的文本进行情感识别：

import pickle
from sklearn.model_selection import train_test_split
from sklearn.metrics import classification_report
from keras.models import Sequential
from keras.layers import Dense, Dropout, Activation, Bidirectional, LSTM, GRU
from keras.callbacks import TensorBoard

def load_f(path):
    with open(path, 'rb')as f:
        data = pickle.load(f)
    return data

path_1 = r'G:/Multimodal/nlp/w2v_weibo_data.pickle'
path_2 = r'G:/Multimodal/labels.pickle'

txts = load_f(path_1)
labels = load_f(path_2)

train_X, test_X, train_Y, test_Y = train_test_split(txts, labels, test_size= 0.2, random_state= 46)

#build model;

tensorboard = TensorBoard(log_dir= r'G:\pytorch')

model = Sequential()
model.add(LSTM(128, input_shape = (None, 100)))
#model.add(GRU(128, input_shape = (None, 100)))
model.add(Dense(3, activation= 'softmax'))
model.compile(loss = 'categorical_crossentropy', optimizer = 'adam', metrics = ['accuracy'])
model.fit(train_X, train_Y, epochs= 50, validation_data = (test_X, test_Y), 
          batch_size= 128, verbose= 1, callbacks= [tensorboard])

y_pre = model.predict(test_X, batch_size= 128)
print(classification_report(test_Y.argmax(axis= 1), y_pre.argmax(axis= 1), digits= 5))

  训练后的结果：

 利用Pytorch对处理好的文本进行情感识别：

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import pickle
from sklearn.model_selection import train_test_split
import torch.utils.data as U
from tqdm import tqdm
from sklearn.metrics import accuracy_score, classification_report

#define Hyper parameters;

Input_size = 100
Hidden_size = 128 
Epochs = 50
batch_size = 128

class lstm(nn.Module):
    def __init__(self):
        super(lstm, self).__init__()
        self.lstm = nn.LSTM(
            input_size=Input_size,
            hidden_size=Hidden_size,
            batch_first=True)
        self.fc = nn.Linear(128, 3)
        
    def forward(self, x):
        out, (h_0, c_0) = self.lstm(x)
        out = out[:, -1, :]
        out = self.fc(out)
        out = F.softmax(out, dim= 1)
        return out, h_0

model = lstm()
optimizer = torch.optim.Adam(model.parameters())

def load_f(path):
    with open(path, 'rb')as f:
        data = pickle.load(f)
    return data

path_1 = r'G:/Multimodal/nlp/w2v_weibo_data.pickle'
path_2 = r'G:/Multimodal/labels.pickle'

txts = load_f(path_1)
labels = load_f(path_2)

train_X, test_X, train_Y, test_Y = train_test_split(txts, labels, test_size= 0.2, random_state= 46)

train_Y = train_Y.argmax(axis=1)
test_Y = test_Y.argmax(axis=1)


train_X = torch.from_numpy(train_X)
train_Y = torch.from_numpy(train_Y)
test_X = torch.from_numpy(test_X)
test_Y = torch.from_numpy(test_Y)


train_data = U.TensorDataset(train_X, train_Y)
test_data = U.TensorDataset(test_X, test_Y)

train_loader = U.DataLoader(train_data, batch_size= batch_size, shuffle=True)
test_loader = U.DataLoader(test_data, batch_size= batch_size, shuffle=False)

def train(model, train_loader, optimizer, epoch):
    model.train()
    for batch_idx, (data, target) in enumerate(train_loader):        
        optimizer.zero_grad()
        output, h_state = model(data)
        labels = output.argmax(dim= 1)
        acc = accuracy_score(target, labels)
        
        #梯度清零；反向传播；
        optimizer.zero_grad()
        loss = F.cross_entropy(output, target)#交叉熵损失函数；
        loss.backward()
        optimizer.step()
        
        
        if (batch_idx+1)%2 == 0:
            finish_rate = (batch_idx* len(data) / len(train_X))*100
            print('Train Epoch: %s'%str(epoch), #Train Epoch: 1
                  '[%d/%d]--'%((batch_idx* len(data)),len(train_X)), #[1450/60000]
                  '%.3f%%'%finish_rate, #0.024
                 '\t', 'acc: %.5f'%acc, #acc: 0.98000
                 '  loss: %s'%loss.item()) #loss: 1.4811499118804932
            
def valid(model, test_loader):
    model.eval()
    test_loss = 0
    acc = 0
    y_true = []
    y_pred = []
    with torch.no_grad():
        for data, target in test_loader:
            output, h_state = model(data)
            test_loss += F.cross_entropy(output, target, reduction='sum').item() # 将一批的损失相加
            output = output.argmax(dim = 1)
            y_true.extend(target)
            y_pred.extend(output)

    acc = accuracy_score(y_true, y_pred)
    #print(classification_report(y_true, y_pred, digits= 5))
    test_loss /= len(test_X)
    
    print('Valid set: Avg Loss:%s'%str(test_loss),
         '\t', 'Avg acc:%s'%str(acc))
    
def test(model, test_loader):
    model.eval()
    y_true = []
    y_pre = []
    for data, target in test_loader:
        output, h_state = model(data)
        output = output.argmax(dim= 1)
        y_true.extend(target)
        y_pre.extend(output)
    print(classification_report(y_true, y_pre, digits= 5))


for epoch in tqdm(range(1, 50)):
    train(model, train_loader, optimizer, epoch)
    valid(model, test_loader)
    print('============================================================================')
print('**********************test set*************************')
test(model, test_loader)

  训练结果：

总结：

   总的来说，利用两个不同的深度学习框架进行训练，两者结果相差并不是很大（要说明的是，两者训练和测试的数据并不是一样的）。但是利用pytorch，可以更好的了解和体会所用的模型。

  在实验过程中，对于pytorch，我有一点很疑惑。同样一个模型，在两个深度学习框架中，所用的配置是完全一样的，在keras中，模型能够快速的拟合，达到一个高精度。但是在pytorch中，确是需要更多的epoch才能拟合。这是我非常困惑的地方，一度认为在模型构建中出了问题，如果有同学知道问题所在，请帮忙解惑，谢谢。