Pytorch学习笔记(4)——从0实现CNN情感分析

感觉之前RNN的代码写的太丑陋了，所以该文章主要参考了Dive-into-DL-PyTorch和中文文本分类 pytorch实现的代码。

1 项目框架

整个项目的框架抽象来看是如下的：

简而言之就是输入的是整一句话，宽度为词向量维度（这里是300维），高度为句子最大长度。经过嵌入层嵌入后，通过卷积与最大池化层，最后进入全连接层，在softmax后得到情感分类的输出。

其详细的框架图为：中文文本分类 pytorch实现

2 预处理

预处理的大部分内容包括数据集与我上篇博客相同：Pytorch学习笔记(3)——从0实现RNN情感分析。只是说将词嵌入这一部分取消了，而是放在的神经网络中进行。更改的部分为：

2.1 将所有词映射为词向量

比起之前将每句话给编码为向量，这里只是将词语与其词向量对应起来。首先，在w2v词向量中已经有了UNK的词向量，所以遇到未登录词，我们可以直接用UNK来表达。但是遇到句子长度不足max_seq_length的时候，我们需要为其填充词向量。我们先定义pad这个变量，其值为。

pad = ''

接着在构建id与词的映射关系时，将词库的最后一维定义为pad。

def get_idx_word_mapping(cleaned_X):
    """
    获得id与词语之间的映射关系
    :param cleaned_X: list
            [[word11, word12, ...], [word21, word22, ...], ...]
            清洗后的文本数据
    :return idx2char: dict
            {1: 'word1', 2: 'word2', ...}
            id与词的映射
    :return char2idx: dict
            {'word1': 1, 'word2': 2, ...}
            词与id的映射
    :return word_set: set
            数据中的所有词
    """
    idx2char = {}
    char2idx = {}
    word_list = []
    # 获得全部词语
    for sentence in cleaned_X:
        for word in sentence:
            word_list.append(word)

    word_set = set(word_list)  # 去重

    # for i, word in enumerate(word_set, start=1):
    for i, word in enumerate(word_set):
        idx2char[i] = word
        char2idx[word] = i

    idx2char[len(word_set)] = pad
    char2idx[pad] = len(word_set)

    return idx2char, char2idx, word_set

于是idx2char与char2idx变为了：

idx2char: {..., last_idx: ''}
char2idx: {..., '': last_idx}

接着根据id与词的映射关系，构建词与词向量之间的映射关系：

def find_word_embedding(w2v, idx2char):
    """
    找到词对应的w2v
    :param w2v: Object
            词向量
    :param idx2char: dict
            id与词的映射
    :return embed: tensor
            id与词向量的映射
    """
    embed = torch.zeros((len(idx2char), 300))
    count = 0
    for id_ in idx2char.keys():
        if id_ == len(idx2char) - 1:
            embed[id_] = torch.zeros(300)
        else:
            try:
                embed[id_] = torch.FloatTensor(w2v[idx2char[id_]])
            except KeyError:
                embed[id_] = torch.FloatTensor(w2v['UNK'])
                count += 1

    print('OOV: ', count)

    return embed

这里要先判断该词语是否是，如果是，则设置为300维全0的向量，否则就判断词语是否是未登录词，如果是未登录词，则设置为UNK的词向量，否则就是原词语的词向量。

2.2 将句子中的词语映射为id

最后，对语料库中的所有句子进行处理，处理的内容包括：

1. 如果词语数不足max_seq_length，则补充至max_seq_length；
2. 如果词语数超出max_seq_length，则在max_seq_length处截断句子；
3. 将句子中的词语转换为id。

def sentence2ids(cleaned_X, char2idx):
    """
    对句子进行处理
        1. 将句子给统一切割或补充到max_seq_length
        2. 将词语替换为id
    :param cleaned_X: list
            句子中的词语
    :param char2idx: dict
            词语与id的映射
    :return: tensor
            句子中每个词语与id之间的映射
    """
    sentences_ids = []

    for sentence in cleaned_X:
        word_count = 0
        sentence_ids = []
        for word in sentence:
            sentence_ids.append(char2idx[word])
            word_count += 1
            # 如果词语数超过了max_seq_length, 则切割
            if word_count >= max_seq_length:
                break
        # 如果词语数不足max_seq_length, 则填补
        if word_count < max_seq_length:
            sentence_ids.extend([char2idx[pad]] * (max_seq_length - word_count))

        sentences_ids.append(sentence_ids)

    return torch.LongTensor(sentences_ids)

3 CNN模型

整个CNN的框架结构可以看博客：中文文本分类 pytorch实现，我也只是照着他的框架结构给实现出来。整体而言，就是先将2.1中生成的embed输入到Embedding层中做映射关系，当输入2.2中的sentences_ids时，根据句子中的id就能查找到对应的词向量。简而言之就是一个lookup function。接着有三个卷积层，其宽为300维，即词向量维度，高分别为[2, 3, 4]，也就是说卷积后的输出的高分别为30-2+1, 30-3+1, 30-4+1，即max_seq_length - filter_size + 1。卷积层的输入只有一个通道，输出有128个通道。其输出通过ReLU做激活函数后，经过全局最大池化层，将每个通道池化后的结果拼接在一起，通过Dropout后，放入全连接层，即可获得最后结果。

class CNN(nn.Module):

    def __init__(self, embed, num_filters, num_outputs, is_freeze=True):
        super().__init__()
        self.filter_sizes = (2, 3, 4)
        self.embeddings = nn.Embedding.from_pretrained(embed, freeze=is_freeze)
        self.convs = nn.ModuleList(
            [nn.Conv2d(1, num_filters, (k, 300)) for k in self.filter_sizes]
        )
        self.relu = nn.ReLU()
        self.dropout = nn.Dropout(0.5)
        self.linear = nn.Linear(num_filters * len(self.filter_sizes), num_outputs)
        self.softmax = nn.Softmax()

    def global_max_pool1d(self, x):
        """
        全局最大池化层, 对高做池化
        :param x: tensor
                shape: (batch_size, output, height)
        :return: tensor
                shape: (batch_size, output, 1)
        """
        return F.max_pool1d(x, kernel_size=x.shape[2])
    
    def forward(self, inputs):
        """
        前向传播
        :param inputs: tensor
                shape: (batch_size, seq_len)
        :return out: tensor
                shape: (batch_size, num_outputs)
        """
        # height_1: seq_len, width_1: w2v_dim
        embed = self.embeddings(inputs)  # shape: (batch_size, height_1, width_1)
        # 添加一维通道维, shape: (batch_size, in_channel, height_1, width_1)
        embed = embed.unsqueeze(1)

        pool_outs = []
        for conv in self.convs:
            # height_2: height_1 - filter_size + 1, width_2 = 1
            # shape: (batch_size, output, height_2, width_2)
            conv_out = conv(embed)
            conv_relu_out = self.relu(conv_out).squeeze(3)  # 清理掉width, 因为是1
            # shape: (batch_size, out, 1)
            pool_out = self.global_max_pool1d(conv_relu_out).squeeze(2)  # 清理掉height, 因为是1
            pool_outs.append(pool_out)
        # shape: (batch_size, out * len(filter_sizes))
        pool_outs = torch.cat(pool_outs, 1)
        pool_dropout_out = self.dropout(pool_outs)
        out = self.linear(pool_dropout_out)
        return self.softmax(out)

训练与测试代码为：

def evaluate(model, data_iter):
    """
    准确率
    :param model: Object
            模型
    :param data_iter: DataLoader
            验证集或测试集
    """
    model.eval()
    loss_total = 0
    predict_all = np.array([], dtype=int)
    labels_all = np.array([], dtype=int)
    with torch.no_grad():
        for texts, labels in data_iter:
            outputs = model(texts)
            loss = F.cross_entropy(outputs, labels)
            loss_total += loss
            labels = labels.data.cpu().numpy()
            predic = torch.max(outputs.data, 1)[1].cpu().numpy()
            labels_all = np.append(labels_all, labels)
            predict_all = np.append(predict_all, predic)

    acc = metrics.accuracy_score(labels_all, predict_all)
    print('test total loss %f, test accuracy %f' % (loss_total / len(data_iter), acc))


def train(X, y, embed, lr, num_epochs):
    """
    训练数据
    :param X: tensor
            shape: (num_sentences, max_seq_length)
            输入数据
    :param y: list
            shape: (num_sentences)
            输入标签
    :param embed: tensor
            shape: (vocab_size, 300)
            词嵌入
    :param lr: float
            学习率
    :param num_epochs: int
            epochs
    """
    batch_size = 8
    y = torch.LongTensor(y)
    X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.1)

    train_dataset = Data.TensorDataset(X_train, y_train)
    test_dataset = Data.TensorDataset(X_test, y_test)
    train_iter = Data.DataLoader(train_dataset, batch_size=batch_size, shuffle=True)
    test_iter = Data.DataLoader(test_dataset, batch_size=batch_size)

    model = CNN(embed, 128, 2)
    loss = nn.CrossEntropyLoss()
    optimizer = torch.optim.Adam(model.parameters(), lr=lr)

    total_iter = 0  # 记录训练了多少轮
    for epoch in range(num_epochs):
        model.train()
        for X, y in train_iter:
            y_hat = model(X)
            optimizer.zero_grad()
            l = loss(y_hat, y)
            l.backward()
            optimizer.step()

            if (total_iter + 1) % 100 == 0:
                pred = torch.max(y_hat.data, 1)[1]  # 哪个位置是预测值
                train_acc = metrics.accuracy_score(y, pred)

                print('iter %d, train_loss %f, train accuracy %f' % (total_iter + 1, l.item(), train_acc))

            total_iter += 1

    evaluate(model, test_iter)

结果为：

iter 10, train_loss 0.622447, train accuracy 0.625000
iter 20, train_loss 0.413261, train accuracy 0.875000
iter 30, train_loss 0.347180, train accuracy 1.000000
iter 40, train_loss 0.314206, train accuracy 1.000000
iter 50, train_loss 0.318095, train accuracy 1.000000
...
iter 1180, train_loss 0.313262, train accuracy 1.000000
iter 1190, train_loss 0.313262, train accuracy 1.000000
iter 1200, train_loss 0.313262, train accuracy 1.000000
test total loss 0.604928, test accuracy 0.727273

由于我这里没有加early stop，也没有加验证集，所以可能会有点过拟合的情况出现。

4 参考

[1] 胡文星. 中文文本分类 pytorch实现[EB/OL]. (2019-08-01)[2021-10-26]. https://zhuanlan.zhihu.com/p/73176084
[2] Aston Zhang, Zachary C. Lipton, Mu Li, Alexander J. Smola. Dive into Deep Learning[M]. http://www.d2l.ai, 2020.