利用BiLSTM网络实现文本分类

这里数据的Preprocess过程不贴代码了，训练过程（train）也不贴了，只是记录一下BiLSTM网络实现代码：

这里隐层数为2，词向量为100维。

import tensorflow as tf
from tensorflow.contrib import rnn


class Model(object):
    def __init__(self, num_layers, seq_length, embedding_size, vocab_size, rnn_size, label_size):
        # 输入数据以及数据标签
        self.input_x = tf.placeholder(tf.int32, [None, seq_length], name="input_x1")
        self.input_y = tf.placeholder(tf.float32, [None, label_size], name="input_y")
        self.dropout_keep_prob = tf.placeholder(tf.float32, name="dropout_keep_prob")
        self.l2_loss = tf.constant(0.0)

        with tf.name_scope('embeddingLayer'):
            # w : 词表（embedding 向量），后面用来训练.
            w = tf.Variable(tf.random_uniform([vocab_size, embedding_size], -1.0, 1.0), name='W')
            embedded = tf.nn.embedding_lookup(w, self.input_x)

            # 根据第二维展开,维度从0开始
            # 删除所有大小为1的维度,删除[1]为要删除维度的参数
            inputs = tf.split(embedded, seq_length, 1)
            inputs = [tf.squeeze(input_, [1]) for input_ in inputs]

        with tf.name_scope("fw"):
            stacked_rnn_fw = []
            for _ in range(num_layers):
                fw_cell = tf.nn.rnn_cell.BasicLSTMCell(rnn_size, forget_bias=1.0, state_is_tuple=True)
                stacked_rnn_fw.append(fw_cell)
            lstm_fw_cell_m = tf.nn.rnn_cell.MultiRNNCell(cells=stacked_rnn_fw, state_is_tuple=True)

        with tf.name_scope("bw"):
            stacked_rnn_bw = []
            for _ in range(num_layers):
                bw_cell = tf.nn.rnn_cell.BasicLSTMCell(rnn_size, forget_bias=1.0, state_is_tuple=True)
                stacked_rnn_bw.append(bw_cell)
            lstm_bw_cell_m = tf.nn.rnn_cell.MultiRNNCell(cells=stacked_rnn_bw, state_is_tuple=True)

        with tf.name_scope("output"):
            outputs, _, _ = rnn.static_bidirectional_rnn(lstm_fw_cell_m, lstm_bw_cell_m, inputs, dtype=tf.float32)

        with tf.name_scope("result"):
            w = tf.Variable(tf.random_uniform([2 * rnn_size, label_size], -1.0, 1.0), name='W')
            b = tf.get_variable('b', [label_size])
            self.output = tf.nn.xw_plus_b(outputs[-1], w, b)
            self.logits = tf.nn.softmax(self.output, dim=1)

        with tf.name_scope("loss"):
            self.losses = tf.nn.softmax_cross_entropy_with_logits(logits=self.logits, labels=self.input_y)
            self.loss = tf.reduce_mean(self.losses)

        with tf.name_scope("accuracy"):
            self.accuracys = tf.equal(tf.argmax(self.logits, axis=1), tf.argmax(self.input_y, axis=1), name="equal")
            self.accuracy = tf.reduce_mean(tf.cast(self.accuracys, "float"), name="accuracy")

        # with tf.name_scope("ans"):
        #     self.ans = tf.subtract(tf.ones_like(self.distance), self.distance, name="temp_sim")
        #     print(self.ans)

# blstm = Model(num_layers=3,
#               seq_length=30,
#               embedding_size=399,
#               vocab_size=120,
#               rnn_size=20,
#               label_size=6)