使用RNN完成IMDB电影评论情感分析

- 任务描述
- 一、环境设置
- 二、数据准备
- 2.1 参数设置
- 2.2 用padding的方式对齐数据
- 2.3 用Dataset与DataLoader加载
- 三、模型配置
- 四、模型训练
- 五、模型评估
- 六、模型预测

任务描述

本示例教程演示如何在IMDB数据集上使用RNN网络完成文本分类的任务。IMDB数据集包含对电影评论进行正向和负向标注的数据，共有25000条文本数据作为训练集，25000条文本数据作为测试集。数据集的官方地址为：IMDB Dataset

在这里插入图片描述

一、环境设置

本示例基于飞桨开源框架2.0版本。

import paddle
import numpy as np
import matplotlib.pyplot as plt
import paddle.nn as nn

print(paddle.__version__)  # 查看当前版本

# cpu/gpu环境选择，在 paddle.set_device() 输入对应运行设备。
device = paddle.set_device('gpu')

2.0.1

二、数据准备

由于IMDB是NLP领域中常见的数据集，飞桨框架将其内置，路径为paddle.text.datasets.Imdb。通过mode参数可以控制训练集与测试集。

print('loading dataset...')
train_dataset = paddle.text.datasets.Imdb(mode='train')
test_dataset = paddle.text.datasets.Imdb(mode='test')
print('loading finished')

构建了训练集与测试集后，可以通过word_idx获取数据集的词表。

word_dict = train_dataset.word_idx  # 获取数据集的词表

# add a pad token to the dict for later padding the sequence
word_dict['<pad>'] = len(word_dict)

for k in list(word_dict)[:5]:
    print("{}:{}".format(k.decode('ASCII'), word_dict[k]))

print("...")

for k in list(word_dict)[-5:]:
    print("{}:{}".format(k if isinstance(k, str) else k.decode('ASCII'), word_dict[k]))

print("totally {} words".format(len(word_dict)))

2.1 参数设置

在这里设置词表大小、embedding大小、batch_size等参数。

vocab_size = len(word_dict) + 1
print(vocab_size)
emb_size = 256
seq_len = 200
batch_size = 32
epochs = 2
pad_id = word_dict['<pad>']

classes = ['negative', 'positive']

# 生成句子列表
def ids_to_str(ids):
    words = []
    for k in ids:
        w = list(word_dict)[k]
        words.append(w if isinstance(w, str) else w.decode('ASCII'))
    return " ".join(words)

2.2 用padding的方式对齐数据

文本数据中，每一句话的长度都是不一样的，为了方便后续的神经网络的计算，通常使用padding的方式对齐数据。

# 读取数据归一化处理
def create_padded_dataset(dataset):
    padded_sents = []
    labels = []
    for batch_id, data in enumerate(dataset):
        sent, label = data[0], data[1]
        padded_sent = np.concatenate([sent[:seq_len], [pad_id] * (seq_len - len(sent))]).astype('int32')
        padded_sents.append(padded_sent)
        labels.append(label)
    return np.array(padded_sents), np.array(labels)

# 对train、test数据进行实例化
train_sents, train_labels = create_padded_dataset(train_dataset)
test_sents, test_labels = create_padded_dataset(test_dataset)

# 查看数据大小及举例内容
print(train_sents.shape)
print(train_labels.shape)
print(test_sents.shape)
print(test_labels.shape)

for sent in train_sents[:3]:
    print(ids_to_str(sent))

2.3 用Dataset与DataLoader加载

将前面准备好的训练集与测试集用Dataset与DataLoader封装后，完成数据的加载。

class IMDBDataset(paddle.io.Dataset):
    '''
    继承paddle.io.Dataset类进行封装数据
    '''
    def __init__(self, sents, labels):
        self.sents = sents
        self.labels = labels
    
    def __getitem__(self, index):
        data = self.sents[index]
        label = self.labels[index]

        return data, label

    def __len__(self):
        return len(self.sents)
    
train_dataset = IMDBDataset(train_sents, train_labels)
test_dataset = IMDBDataset(test_sents, test_labels)

train_loader = paddle.io.DataLoader(train_dataset, return_list=True,
                                    shuffle=True, batch_size=batch_size, drop_last=True)
test_loader = paddle.io.DataLoader(test_dataset, return_list=True,
                                    shuffle=True, batch_size=batch_size, drop_last=True)

三、模型配置

本示例中使用一个序列特性的RNN网络，在查找到每个词对应的embedding后，取平均作为一个句子的表示。然后用Linear进行线性变换，同时使用Dropout防止过拟合。

class MyRNN(paddle.nn.Layer):
    def __init__(self):
        super(MyRNN, self).__init__()
        self.embedding = nn.Embedding(vocab_size, 256)
        self.rnn = nn.SimpleRNN(256, 256, num_layers=2, direction='forward',dropout=0.5)
        self.linear = nn.Linear(in_features=256*2, out_features=2)
        self.dropout = nn.Dropout(0.5)
    
    def forward(self, inputs):
        emb = self.dropout(self.embedding(inputs))
        output, hidden = self.rnn(emb)
        hidden = paddle.concat((hidden[-2,:,:], hidden[-1,:,:]), axis = 1)
        hidden = self.dropout(hidden)
        return self.linear(hidden)

四、模型训练

# 可视化定义
def draw_process(title, color, iters, data, label):
    plt.title(title, fontsize=24)
    plt.xlabel("iter", fontsize=20)
    plt.ylabel(label, fontsize=20)
    plt.plot(iters, data, color=color, label=label) 
    plt.legend()
    plt.grid()
    plt.show()

# 对模型进行封装
def train(model):
    model.train()
    opt = paddle.optimizer.Adam(learning_rate=0.001, parameters=model.parameters())
    steps = 0
    Iters, total_loss, total_acc = [], [], []

    for epoch in range(epochs):
        for batch_id, data in enumerate(train_loader):
            steps +=

 1
            sent = data[0]
            label = data[1]
            
            logits = model(sent)
            loss = paddle.nn.functional.cross_entropy(logits, label)
            acc = paddle.metric.accuracy(logits, label)

            if batch_id % 500 == 0:  # 500个epoch输出一次结果
                Iters.append(steps)
                total_loss.append(loss.numpy()[0])
                total_acc.append(acc.numpy()[0])

                print("epoch: {}, batch_id: {}, loss is: {}".format(epoch, batch_id, loss.numpy()))
            
            loss.backward()
            opt.step()
            opt.clear_grad()

        # evaluate model after one epoch
        model.eval()
        accuracies = []
        losses = []
        
        for batch_id, data in enumerate(test_loader):
            sent = data[0]
            label = data[1]

            logits = model(sent)
            loss = paddle.nn.functional.cross_entropy(logits, label)
            acc = paddle.metric.accuracy(logits, label)
            
            accuracies.append(acc.numpy())
            losses.append(loss.numpy())
        
        avg_acc, avg_loss = np.mean(accuracies), np.mean(losses)

        print("[validation] accuracy: {}, loss: {}".format(avg_acc, avg_loss))
        
        model.train()

        # 保存模型
        paddle.save(model.state_dict(), str(epoch) + "_model_final.pdparams")
    
    # 可视化查看
    draw_process("training loss", "red", Iters, total_loss, "training loss")
    draw_process("training acc", "green", Iters, total_acc, "training acc")
        
model = MyRNN()
train(model)

五、模型评估

model_state_dict = paddle.load('1_model_final.pdparams')  # 导入模型
model = MyRNN()
model.set_state_dict(model_state_dict) 
model.eval()
accuracies = []
losses = []

for batch_id, data in enumerate(test_loader):
    sent = data[0]
    label = data[1]

    logits = model(sent)
    loss = paddle.nn.functional.cross_entropy(logits, label)
    acc = paddle.metric.accuracy(logits, label)
    
    accuracies.append(acc.numpy())
    losses.append(loss.numpy())

avg_acc, avg_loss = np.mean(accuracies), np.mean(losses)
print("[validation] accuracy: {}, loss: {}".format(avg_acc, avg_loss))

六、模型预测

def ids_to_str(ids):
    words = []
    for k in ids:
        w = list(word_dict)[k]
        words.append(w if isinstance(w, str) else w.decode('UTF-8'))
    return " ".join(words)

label_map = {0: "negative", 1: "positive"}

# 导入模型
model_state_dict = paddle.load('1_model_final.pdparams')
model = MyRNN()
model.set_state_dict(model_state_dict) 
model.eval()

for batch_id, data in enumerate(test_loader):
    sent = data[0]
    results = model(sent)

    predictions = []
    for probs in results:
        # 映射分类label
        idx = np.argmax(probs)
        labels = label_map[idx]
        predictions.append(labels)
    
    for i, pre in enumerate(predictions):
        print(' 数据: {} \n 情感: {}'.format(ids_to_str(sent[0]), pre))
        break
    break