0x00 前言
通常我们处理双向LSTM的时候,序列长短不一参差不齐,就不可避免的需要进行padding,
实际上句子长度可能如下所示,为[2, 4, 5],就需要被padding成这样才能作为矩阵传入计算:
array([[1, 2, 0, 0, 0],
[2, 1, 4, 3, 0],
[3, 2, 1, 5, 4]], dtype=int32)
对于padding的数据做双向LSTM的时候,反向的部分会将0
也算入,
虽说对于大多任务中,把padding也算作一个词,做一个padding的embedding,
也是可以获得不错的效果的,但是存在过长的句子时,
大多较短句子后面过多的padding会把末状态过于稀释,
诸如此类,还有各种各样的情况都希望可以双向LSTM可以每个方向都只计算有效长度
0x01 简要解释
这里的方法是采用单向LSTM,配以cell_len来控制终止位置,
正向LSTM正常调用即可,反向LSTM需要对非Padding位进行reverse,
(即有效长度部分逆转,Padding部分不变)
这里和通常直接调用的双向LSTM不同,自定义的成分比较多,例如下面实际场景运用中,
正反向错位合并,普通的LSTM应该不会那么轻松(有点说大话了,至少……看起来没那么明显?)
0x02 函数介绍
# Tensorflow.Reverse()
sess = tf.Session()
inp = tf.placeholder(tf.int32, [None, 5])
cell_lens = tf.placeholder(tf.int32, [None])
rseq = tf.reverse_sequence(
inp, cell_lens, # 输入矩阵 与 需要作reverse操作的长度
seq_axis=None,
batch_axis=None,
name='reverse_data', # 取名,方便在计算图中定位
seq_dim=1, batch_dim=0 # 需要作reverse操作的维度和作为batch的维度
)
sess.run(rseq, {
inp: [range(5), range(5), range(5)],
cell_lens: [2, 3, 4]
})
"""
# 分别前2、3、4维进行reverse操作
array([[1, 0, 2, 3, 4],
[2, 1, 0, 3, 4],
[3, 2, 1, 0, 4]], dtype=int32)
"""
0x03 实际场景运用节选
这是在一处NLP相关使用场景中运用的案例
由于研究相关及保密需求,此处省略数据流部分,仅展示与本文讨论相关要点
class LstmLayer(object):
""" LSTM layer class """
def __init__(self, num_units, bidirection=False, sequence_length=None, name="lstm"):
self.num_units = num_units
self.bidirection = bidirection
self.sequence_length = tf.reshape(sequence_length, [-1])
self.name = name
def __call__(self, inputs, time_major=False):
with tf.name_scope('{}_cal'.format(self.name)):
with tf.variable_scope(self.name, reuse=tf.AUTO_REUSE):
if self.bidirection:
lstm_fw_cell = tf.nn.rnn_cell.BasicLSTMCell(self.num_units, state_is_tuple=True)
lstm_bw_cell = tf.nn.rnn_cell.BasicLSTMCell(self.num_units, state_is_tuple=True)
outputs, output_states = tf.nn.bidirectional_dynamic_rnn(
lstm_fw_cell, lstm_bw_cell, inputs,
sequence_length=self.sequence_length,
time_major=time_major, dtype=tf.float32)
else:
lstm_fw_cell = tf.nn.rnn_cell.BasicLSTMCell(self.num_units, state_is_tuple=True)
outputs, output_states = tf.nn.dynamic_rnn(
lstm_fw_cell, inputs, sequence_length=self.sequence_length,
time_major=time_major, dtype=tf.float32)
return outputs, output_states
class Network(object):
""" Network for model testing """
def __init__(self, options=feed_options, is_training=True):
self.options = options
self.is_training = is_training
# LSTM Layer
self.forward_lstm = LstmLayer(num_units=options.get('lstm_dim'),
sequence_length=self.cell_lens,
name="forward_lstm")
self.backward_lstm = LstmLayer(num_units=options.get('lstm_dim'),
sequence_length=self.cell_lens,
name="backward_lstm")
def get_reverse(self, input, seq_dim=1, batch_dim=0):
"""
# Given this:
batch_dim = 2
seq_dim = 0
input.dims = (8, ?, 4, ...)
seq_lengths = [7, 2, 3, 5]
# then slices of input are reversed on seq_dim, but only up to seq_lengths:
output[0:7, :, 0, :, ...] = input[7:0:-1, :, 0, :, ...]
output[0:2, :, 1, :, ...] = input[2:0:-1, :, 1, :, ...]
output[0:3, :, 2, :, ...] = input[3:0:-1, :, 2, :, ...]
output[0:5, :, 3, :, ...] = input[5:0:-1, :, 3, :, ...]
# while entries past seq_lens are copied through:
output[7:, :, 0, :, ...] = input[7:, :, 0, :, ...]
output[2:, :, 1, :, ...] = input[2:, :, 1, :, ...]
output[3:, :, 2, :, ...] = input[3:, :, 2, :, ...]
output[2:, :, 3, :, ...] = input[2:, :, 3, :, ...]
"""
tf.reverse_sequence(
input, self.cell_lens,
name='reverse_data',
seq_dim=seq_dim,
batch_dim=batch_dim,
)
def lstm_layer(self, forward_emb):
# [batch, seg_len + 2, emb_dim] -> [batch, seg_len + 2, emb_dim]
backward_emb = self.get_reverse(forward_emb)
# [batch, seg_len + 2, emb_dim] x 2 -> [batch, seg_len, lstm_dim] x 2
_, (c1, f_hidden) = self.forward_lstm(forward_emb)
_, (c2, b_hidden) = self.backward_lstm(backward_emb)
# [batch, seg_len, lstm_dim] x 2 -> [batch, seg_len, lstm_dim * 2]
return tf.concat([f_hidden, self.get_reverse(b_hidden)], -1)
此处应用场景中还有卖弄了一个反向错位相加的小伎俩,使用了tensorflow.reverse()
,
拿到需求就想出可以通过前后加<BOS>
、<EOS>
后,正反向各少两个节点做LSTM隐向量拼接这一点,
我自己还是有点小骄傲的~~