支招 | 使用Pytorch进行文本分类

AI研习社

发布于 2019-09-17 17:04:34

2.2K0

发布于 2019-09-17 17:04:34

文章被收录于专栏：AI研习社

本文为 AI 研习社社区用户 @Dendi 的博客文章，欢迎扫描底部社区名片访问 @Dendi 的主页，查看更多内容。

01. BILSTM+ATTENTION

网络结构

代码实现

Attention计算

将BILSTM网络输出的结果（shape：[batch_size, time_step, hidden_dims * num_directions(=2)]）拆成两个大小为[batch_size, time_step, hidden_dims]的Tensor；
将第一步拆出的两个Tensor进行相加运算得到h（shape：[batch_size, time_step, hidden_dims]）；
将h作为输入，通过self.attention_layer得到attention的计算向量atten_w（shape：[batch_size, time_step, hidden_dims]）；
将第二步的h进行tanh()激活，得到m（shape：[batch_size, time_step, hidden_dims]），留待后续进行残差计算；
将atten_w的2、3维度进行调换，并与m进行矩阵的乘法运算，得到atten_context（shape：[batch_size, time_step, time_step]）;
将atten_context最后一维进行softmax计算，得到当前时刻相对于所有时刻的权重：softmax_w（shape：[batch_size, time_step, time_step]）；
将h的2、3维度进行调换，并与softmax_w进行矩阵运算，得到基于权重的context（shape：[batch_size, hidden_dims, time_step]）；
将h的2、3维度进行调换，并与context进行求和运算，得到context_with_attn（shape：[batch_size, hidden_dims, time_step]）；
将context_with_attn的最后一维进行求和，得到result（shape：[batch_size, hidden_dims]）；
计算结束，返回result；

模型效果

1层BILSTM在训练集准确率：99.8%，测试集准确率：96.5%；

2层BILSTM在训练集准确率：99.9%，测试集准确率：97.3%；

调参

dropout的值要在 0.1 以下（经验之谈，笔者在实践中发现，dropout取0.1时比dropout取0.3时在测试集准确率能提高0.5%）。

02. Transformer

前言

文本分类不是生成式的任务，因此只使用Transformer的编码部分（Encoder）进行特征提取。如果不熟悉Transformer模型的原理请移步：https://blog.csdn.net/dendi_hust/article/details/98759771

网络结构

代码实现

自注意力模型：

class TextSlfAttnNet(nn.Module):
    ''' 自注意力模型 '''

    def __init__(self,
                 config: TextSlfAttnConfig,
                 char_size=5000,
                 pinyin_size=5000):
        super(TextSlfAttnNet, self).__init__()
        # 字向量
        self.char_embedding = nn.Embedding(char_size, config.embedding_size)
        # 拼音向量
        self.pinyin_embedding = nn.Embedding(pinyin_size, config.embedding_size)
        # 位置向量
        self.pos_embedding = nn.Embedding.from_pretrained(
            get_sinusoid_encoding_table(config.max_sen_len, config.embedding_size, padding_idx=0),
            freeze=True)

        self.layer_stack = nn.ModuleList([
            EncoderLayer(config.embedding_size, config.hidden_dims, config.n_heads, config.k_dims, config.v_dims, dropout=config.keep_dropout)
            for _ in range(config.hidden_layers)
        ])

        self.fc_out = nn.Sequential(
            nn.Dropout(config.keep_dropout),
            nn.Linear(config.embedding_size, config.hidden_dims),
            nn.ReLU(inplace=True),
            nn.Dropout(config.keep_dropout),
            nn.Linear(config.hidden_dims, config.num_classes),

        )


    def forward(self, char_id, pinyin_id, pos_id):
        char_inputs = self.char_embedding(char_id)
        pinyin_iputs = self.pinyin_embedding(pinyin_id)

        sen_inputs = torch.cat((char_inputs, pinyin_iputs), dim=1)
        # sentence_length = sen_inputs.size()[1]
        # pos_id = torch.LongTensor(np.array([i for i in range(sentence_length)]))
        pos_inputs = self.pos_embedding(pos_id)
        # batch_size * sen_len * embedding_size
        inputs = sen_inputs + pos_inputs

        for layer in self.layer_stack:
            inputs, _ = layer(inputs)

        enc_outs = inputs.permute(0, 2, 1)
        enc_outs = torch.sum(enc_outs, dim=-1)
        return self.fc_out(enc_outs)

编码层（EncoderLayer）


class EncoderLayer(nn.Module):
    '''编码层'''

    def __init__(self, d_model, d_inner, n_head, d_k, d_v, dropout=0.1):
        '''

        :param d_model: 模型输入维度
        :param d_inner: 前馈神经网络隐层维度
        :param n_head:  多头注意力
        :param d_k:     键向量
        :param d_v:     值向量
        :param dropout:
        '''
        super(EncoderLayer, self).__init__()
        self.slf_attn = MultiHeadAttention(n_head, d_model, d_k, d_v, dropout=dropout)
        self.pos_ffn = PositionwiseFeedForward(d_model, d_inner, dropout=dropout)

    def forward(self, enc_input, non_pad_mask=None, slf_attn_mask=None):
        '''

        :param enc_input:
        :param non_pad_mask:
        :param slf_attn_mask:
        :return:
        '''
        enc_output, enc_slf_attn = self.slf_attn(enc_input, enc_input, enc_input, mask=slf_attn_mask)
        if non_pad_mask is not None:
            enc_output *= non_pad_mask

        enc_output = self.pos_ffn(enc_output)
        if non_pad_mask is not None:
            enc_output *= non_pad_mask
        return enc_output, enc_slf_attn

多头注意力（`MultiHeadAttention`）

class MultiHeadAttention(nn.Module):
    '''
        “多头”注意力模型
    '''

    def __init__(self, n_head, d_model, d_k, d_v, dropout=0.1):
        '''

        :param n_head: “头”数
        :param d_model: 输入维度
        :param d_k: 键向量维度
        :param d_v: 值向量维度
        :param dropout:
        '''
        super(MultiHeadAttention, self).__init__()

        self.n_head = n_head
        self.d_k = d_k
        self.d_v = d_v
        # 产生 查询向量q，键向量k， 值向量v
        self.w_qs = nn.Linear(d_model, n_head * d_k)
        self.w_ks = nn.Linear(d_model, n_head * d_k)
        self.w_vs = nn.Linear(d_model, n_head * d_v)

        nn.init.normal_(self.w_qs.weight, mean=0, std=np.sqrt(2.0 / (d_model + d_k)))
        nn.init.normal_(self.w_ks.weight, mean=0, std=np.sqrt(2.0 / (d_model + d_k)))
        nn.init.normal_(self.w_vs.weight, mean=0, std=np.sqrt(2.0 / (d_model + d_v)))

        self.attention = ScaledDotProductAttention(temperature=np.power(d_k, 0.5))
        self.layer_normal = nn.LayerNorm(d_model)

        self.fc = nn.Linear(n_head * d_v, d_model)
        nn.init.xavier_normal_(self.fc.weight)

        self.dropout = nn.Dropout(dropout)
        
      def forward(self, q, k, v, mask=None):
        '''
        计算多头注意力
        :param q: 用于产生  查询向量
        :param k: 用于产生  键向量
        :param v:  用于产生 值向量
        :param mask:
        :return:
        '''
        d_k, d_v, n_head = self.d_k, self.d_v, self.n_head

        sz_b, len_q, _ = q.size()
        sz_b, len_k, _ = k.size()
        sz_b, len_v, _ = v.size()

        residual = q

        q = self.w_qs(q).view(sz_b, len_q, n_head, d_k)
        k = self.w_ks(k).view(sz_b, len_k, n_head, d_k)
        v = self.w_vs(v).view(sz_b, len_v, n_head, d_v)

        # (n*b) x lq x dk
        q = q.permute(2, 0, 1, 3).contiguous().view(-1, len_q, d_k)
        # (n*b) x lk x dk
        k = k.permute(2, 0, 1, 3).contiguous().view(-1, len_k, d_k)
        # (n*b) x lv x dv
        v = v.permute(2, 0, 1, 3).contiguous().view(-1, len_v, d_v)

        # mask = mask.repeat(n_head, 1, 1)  # (n*b) x .. x ..
        #
        output, attn = self.attention(q, k, v, mask=None)
        # (n_heads * batch_size) * lq * dv
        output = output.view(n_head, sz_b, len_q, d_v)
        # batch_size * len_q * (n_heads * dv)
        output = output.permute(1, 2, 0, 3).contiguous().view(sz_b, len_q, -1)
        output = self.dropout(self.fc(output))
        output = self.layer_normal(output + residual)
        return output, attn

前馈神经网络（`PositionwiseFeedForward`）

class PositionwiseFeedForward(nn.Module):
    '''
        前馈神经网络
    '''

    def __init__(self, d_in, d_hid, dropout=0.1):
        '''

        :param d_in:    输入维度
        :param d_hid:   隐藏层维度
        :param dropout:
        '''
        super(PositionwiseFeedForward, self).__init__()
        self.w_1 = nn.Conv1d(d_in, d_hid, 1)
        self.w_2 = nn.Conv1d(d_hid, d_in, 1)
        self.layer_normal = nn.LayerNorm(d_in)
        self.dropout = nn.Dropout(dropout)

    def forward(self, x):
        residual = x
        output = x.transpose(1, 2)
        output = self.w_2(F.relu(self.w_1(output)))
        output = output.transpose(1, 2)
        output = self.dropout(output)
        output = self.layer_normal(output + residual)
        return output

位置函数

def get_sinusoid_encoding_table(n_position, d_hid, padding_idx=None):
    '''
    计算位置向量
    :param n_position:      位置的最大值
    :param d_hid:           位置向量的维度，和字向量维度相同（要相加求和）
    :param padding_idx: 
    :return: 
    '''

    def cal_angle(position, hid_idx):
        return position / np.power(10000, 2 * (hid_idx // 2) / d_hid)

    def get_posi_angle_vec(position):
        return [cal_angle(position, hid_j) for hid_j in range(d_hid)]

    sinusoid_table = np.array([get_posi_angle_vec(pos_i) for pos_i in range(n_position)])

    sinusoid_table[:, 0::2] = np.sin(sinusoid_table[:, 0::2])  # dim 2i
    sinusoid_table[:, 1::2] = np.cos(sinusoid_table[:, 1::2])  # dim 2i+1

    if padding_idx is not None:
        # zero vector for padding dimension
        sinusoid_table[padding_idx] = 0.

    return torch.FloatTensor(sinusoid_table)

实践经验

在分类任务中，与`BILSTM+ATTENTION`相比：

1. 模型比`LSTM`大很多，同样的任务`LSTM`模型6M左右，`Transformer`模型55M；

2. 收敛速度比较慢；

3. 超参比较多，不易调参，但同时也意味着弹性比较大；

4. 效果和`BILSTM`模型差不多。

03. TextCNN

网络结构

代码实现

class TextCNN(nn.Module):


def __init__(self,
                 config:TCNNConfig,
                 char_size = 5000, pinyin_size=5000):
        super(TextCNN, self).__init__()
        self.learning_rate = config.learning_rate
        self.keep_dropout = config.keep_dropout
        self.sequence_length = config.sequence_length
        self.char_embedding_size = config.char_embedding_size
        self.pinyin_embedding_size = config.pinyin_embedding_size
        self.filter_list = config.filter_list
        self.out_channels = config.out_channels
        self.l2_reg_lambda = config.l2_reg_lambda
        self.model_dir = config.model_dir
        self.data_save_frequency = config.data_save_frequency
        self.model_save_frequency = config.model_save_frequency
        self.char_size = char_size
        self.pinyin_size = pinyin_size
        self.embedding_size = self.char_embedding_size
        self.total_filters_size = self.out_channels * len(self.filter_list)
        self.build_model()


def build_model(self):
# 初始化字向量
        self.char_embeddings = nn.Embedding(self.char_size, self.char_embedding_size)
# 字向量参与更新
        self.char_embeddings.weight.requires_grad = True
# 初始化拼音向量
        self.pinyin_embeddings = nn.Embedding(self.pinyin_size, self.pinyin_embedding_size)
        self.pinyin_embeddings.weight.requires_grad = True
        self.conv_list = nn.ModuleList()


        conv_list = [nn.Sequential(
            nn.Conv1d(self.embedding_size, self.out_channels, filter_size),
            nn.BatchNorm1d(self.out_channels),
            nn.ReLU(inplace=True)
) for filter_size in self.filter_list]
# 卷积列表
        self.conv_lists_layer = nn.ModuleList(conv_list)






        self.output_layer = nn.Sequential(
            nn.Dropout(self.keep_dropout),
            nn.Linear(self.total_filters_size, self.total_filters_size),
            nn.ReLU(inplace=True),
            nn.Linear(self.total_filters_size, 2)
)


def forward(self, char_id, pinyin_id):
# char_id = torch.from_numpy(np.array(input[0])).long()
# pinyin_id = torch.from_numpy(np.array(input[1])).long()
        pooled_outputs = []
        sen_char = self.char_embeddings(char_id)
        sen_pinyin = self.pinyin_embeddings(pinyin_id)
        sen_embed = torch.cat((sen_char, sen_pinyin), dim=1)
# 转换成 (N C SEN_LEN) 的形式
        sen_embed = sen_embed.permute(0, 2, 1)
for conv in self.conv_lists_layer:
# print(sen_embed.shape)
            conv_output = conv(sen_embed)
            max_polling_output = torch.max(conv_output, dim=2)
            pooled_outputs.append(max_polling_output[0])


        total_pool = torch.cat(pooled_outputs, 1)
        flatten_pool = total_pool.view(-1, self.total_filters_size)
        fc_output = self.output_layer(flatten_pool)
return fc_output

本文参与腾讯云自媒体同步曝光计划，分享自微信公众号。

原始发表：2019-09-12，如有侵权请联系 cloudcommunity@tencent.com 删除

机器学习