任務(wù)描述

本示例教程演示如何在IMDB數(shù)據(jù)集上使用RNN網(wǎng)絡(luò)完成文本分類的任務(wù)。IMDB數(shù)據(jù)集包含對電影評論進(jìn)行正向和負(fù)向標(biāo)注的數(shù)據(jù)，共有25000條文本數(shù)據(jù)作為訓(xùn)練集，25000條文本數(shù)據(jù)作為測試集。數(shù)據(jù)集的官方地址為：IMDB Dataset

一、環(huán)境設(shè)置

本示例基于飛槳開源框架2.0版本。

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

print(paddle.__version__)  # 查看當(dāng)前版本

# cpu/gpu環(huán)境選擇，在 paddle.set_device() 輸入對應(yīng)運(yùn)行設(shè)備。
device = paddle.set_device('gpu')

2.0.1

二、數(shù)據(jù)準(zhǔn)備

由于IMDB是NLP領(lǐng)域中常見的數(shù)據(jù)集，飛槳框架將其內(nèi)置，路徑為paddle.text.datasets.Imdb。通過mode參數(shù)可以控制訓(xùn)練集與測試集。

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

構(gòu)建了訓(xùn)練集與測試集后，可以通過word_idx獲取數(shù)據(jù)集的詞表。

word_dict = train_dataset.word_idx  # 獲取數(shù)據(jù)集的詞表

# 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 參數(shù)設(shè)置

在這里設(shè)置詞表大小、embedding大小、batch_size等參數(shù)。

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的方式對齊數(shù)據(jù)

文本數(shù)據(jù)中，每一句話的長度都是不一樣的，為了方便后續(xù)的神經(jīng)網(wǎng)絡(luò)的計算，通常使用padding的方式對齊數(shù)據(jù)。

# 讀取數(shù)據(jù)歸一化處理
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)

# 對train、test數(shù)據(jù)進(jìn)行實(shí)例化
train_sents, train_labels = create_padded_dataset(train_dataset)
test_sents, test_labels = create_padded_dataset(test_dataset)

# 查看數(shù)據(jù)大小及舉例內(nèi)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與DataLoader加載

將前面準(zhǔn)備好的訓(xùn)練集與測試集用Dataset與DataLoader封裝后，完成數(shù)據(jù)的加載。

class IMDBDataset(paddle.io.Dataset):
    '''
    繼承paddle.io.Dataset類進(jìn)行封裝數(shù)據(jù)
    '''
    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

本示例中使用一個序列特性的RNN網(wǎng)絡(luò)，在查找到每個詞對應(yīng)的embedding后，取平均作為一個句子的表示。然后用Linear進(jìn)行線性變換，同時使用Dropout防止過擬合。

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、模型訓(xùn)練

# 可視化定義
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()

# 對模型進(jìn)行封裝
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個epoch輸出一次結(jié)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")
    
    # 可視化查看
    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、模型評估

model_state_dict = paddle.load('1_model_final.pdparams')  # 導(dǎo)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、模型預(yù)測

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"}

# 導(dǎo)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:
        # 映射分類label
        idx = np.argmax(probs)
        labels = label_map[idx]
        predictions.append(labels)
    
    for i, pre in enumerate(predictions):
        print(' 數(shù)據(jù): {} \n 情感: {}'.format(ids_to_str(sent[0]), pre))
        break
    break

以上是使用RNN完成IMDB電影評論情感分析的示例。通過搭建RNN網(wǎng)絡(luò)，對文本數(shù)據(jù)進(jìn)行預(yù)處理、模型訓(xùn)練和評估，最終實(shí)現(xiàn)了對電影評論情感的分類。在實(shí)際應(yīng)用中，可以根據(jù)需求調(diào)整網(wǎng)絡(luò)結(jié)構(gòu)和超參數(shù)，提高模型性能。文章來源地址http://www.zghlxwxcb.cn/news/detail-795410.html

到了這里，關(guān)于使用RNN完成IMDB電影評論情感分析的文章就介紹完了。如果您還想了解更多內(nèi)容，請?jiān)谟疑辖撬阉鱐OY模板網(wǎng)以前的文章或繼續(xù)瀏覽下面的相關(guān)文章，希望大家以后多多支持TOY模板網(wǎng)！