考虑 state_is_tuple

   

Output, new_state = cell(input, state)

   

state其实是两个 一个 c state，一个m(对应下图的hidden 或者h) 其中m(hidden)其实也就是输出

   

   

   

   

new_state = (LSTMStateTuple(c, m) if self._state_is_tuple

else array_ops.concat(1, [c, m]))

return m, new_state

   

   

def basic_rnn_seq2seq(

encoder_inputs, decoder_inputs, cell, dtype=dtypes.float32, scope=None):

with variable_scope.variable_scope(scope or "basic_rnn_seq2seq"):

_, enc_state = rnn.rnn(cell, encoder_inputs, dtype=dtype)

return rnn_decoder(decoder_inputs, enc_state, cell)

   

   

def rnn_decoder(decoder_inputs, initial_state, cell, loop_function=None,

scope=None):

with variable_scope.variable_scope(scope or "rnn_decoder"):

state = initial_state

outputs = []

prev = None

for i, inp in enumerate(decoder_inputs):

if loop_function is not None and prev is not None:

with variable_scope.variable_scope("loop_function", reuse=True):

inp = loop_function(prev, i)

if i > 0:

variable_scope.get_variable_scope().reuse_variables()

output, state = cell(inp, state)

outputs.append(output)

if loop_function is not None:

prev = output

return outputs, state

   

   

这里decoder用了encoder的最后一个state 作为输入

   

然后输出结果是decoder过程最后的state 加上所有ouput的集合(也就是hidden的集合)

注意ouputs[-1]其实数值和state里面的m是一致的

当然有可能后面outputs 用dynamic rnn 会补0

   

encode_feature, state = melt.rnn.encode(

cell,

inputs,

seq_length,

encode_method=0,

output_method=3)

   

encode_feature.eval()

array([[[ 4.27834410e-03, 1.45841937e-03, 1.25767402e-02,

5.00775501e-03],

[ 6.24437723e-03, 2.60074623e-03, 2.32168660e-02,

9.47457738e-03],

[ 7.59789022e-03, -5.34060055e-05, 1.64511874e-02,

-5.71310846e-03],

[ 0.00000000e+00, 0.00000000e+00, 0.00000000e+00,

0.00000000e+00]]], dtype=float32)

   

   

state[1].eval()

array([[ 7.59789022e-03, -5.34060055e-05, 1.64511874e-02,

-5.71310846e-03]], dtype=float32)