新版seqseq接口说明

attention_mechanism = tf.contrib.seq2seq.BahdanauAttention(num_units=FLAGS.rnn_hidden_size, memory = encoder_outputs, memory_sequence_length = encoder_sequence_length)

这一步创造一个attention_mechanism。通过__call__(self, query, previous_alignments）来调用，输入query也就是decode hidden，输入previous_alignments是encode hidden，输出是一个attention概率矩阵

helper = tf.contrib.seq2seq.ScheduledEmbeddingTrainingHelper(inputs, tf.to_int32(sequence_length), emb, tf.constant(FLAGS.scheduled_sampling_probability))

创建一个helper，用来处理每个时刻的输入和输出

my_decoder = tf.contrib.seq2seq.BasicDecoder(cell = cell, helper = helper, initial_state = state)

调用的核心部分。通过def step(self, time, inputs, state, name=None)来控制每一个进行decode

首先把inputs和attention进行concat作为输入。（为什么这样做，参考LSTM的实现 W1U+W2V，其实是把U，V concat在乘以一个W），那么这里inputs就是U，attention就是V（其实tf.concat(query,attention矩阵 * memory）在做个outpreject）。

outputs, state, final_sequence_lengths = tf.contrib.seq2seq.dynamic_decode(my_decoder， scope='seq_decode'）

最后通过dynamic_decode来控制整个flow

写到前面：

先看：

class BasicRNNCell(RNNCell):

def call(self, inputs, state):
"""Most basic RNN: output = new_state = act(W * input + U * state + B)."""
if self._linear is None:
self._linear = _Linear([inputs, state], self._num_units, True)

这个是核心，也就是W * input + U * state + B的实现，tf是用_Linear来实现的（_Linear的实现就是把input和state进行concat，然后乘以一个W）。由于rnn只有hidden，所以这里的state就是hidden

再看

class BasicLSTMCell(RNNCell):

if self._state_is_tuple:
c, h = state
else:
c, h = array_ops.split(value=state, num_or_size_splits=2, axis=1)

if self._linear is None:
self._linear = _Linear([inputs, h], 4 * self._num_units, True)
# i = input_gate, j = new_input, f = forget_gate, o = output_gate
i, j, f, o = array_ops.split(
value=self._linear([inputs, h]), num_or_size_splits=4, axis=1)

new_c = (
c * sigmoid(f + self._forget_bias) + sigmoid(i) * self._activation(j))
new_h = self._activation(new_c) * sigmoid(o)

if self._state_is_tuple:
new_state = LSTMStateTuple(new_c, new_h)
else:
new_state = array_ops.concat([new_c, new_h], 1)
return new_h, new_state

就非常明显了，由于lstm的state是由两部分构成的，一个是hidden，一个是state，第一步先split。之后用inputs和h进行linear，由于我们要输出4个结果，记得输出维度一定要是4*_num_units。然后根据公式再进行后面的操作，最后返回新的hidden和state，也很直观。

之后再看，加入attention之后怎么弄：

我们这里的attention为encode hidden，那么根据公式是attention和decode hidden进行concat作为一个大的hidden，之后和inputs一起进入网络。

但是，tf实现的时候是这样子的，首先把attention和inputs进行concat，之后把连接的结果作为inputs和decode hidden一起送入网络。为什么能这么做呢，是因为在网络内部其实也是concat之后再linear，参考上面的BasicLSTMCell实现，所有关键就是把（inputs，attention，decode hidden）concat一起就行了，不管顺序是啥。说道这里你终于明白了AttentionWrapper到底是干啥的了。那么attention怎么计算呢，有个_compute_attention函数。我感觉就是非常直接了，attention_mechanism是你需要的attention映射矩阵的方式，

def _compute_attention(attention_mechanism, cell_output, previous_alignments,
attention_layer):
"""Computes the attention and alignments for a given attention_mechanism."""
alignments = attention_mechanism(
cell_output, previous_alignments=previous_alignments)

# Reshape from [batch_size, memory_time] to [batch_size, 1, memory_time]
expanded_alignments = array_ops.expand_dims(alignments, 1)
# Context is the inner product of alignments and values along the
# memory time dimension.
# alignments shape is
# [batch_size, 1, memory_time]
# attention_mechanism.values shape is
# [batch_size, memory_time, memory_size]
# the batched matmul is over memory_time, so the output shape is
# [batch_size, 1, memory_size].
# we then squeeze out the singleton dim.
context = math_ops.matmul(expanded_alignments, attention_mechanism.values)
context = array_ops.squeeze(context, [1])

if attention_layer is not None:
attention = attention_layer(array_ops.concat([cell_output, context], 1))
else:
attention = context

return attention, alignments