LSTM在keras中参数return_sequences、return_state的超详细区别(附代码)

一、定义

return_sequences:默认为false。当为false时，返回最后一层最后一个步长的hidden state;当为true时，返回最后一层的所有hidden state。

return_state:默认false.当为true时，返回最后一层的最后一个步长的输出hidden state和输入cell state。

二、实例验证

下图的输入是一个步长为3,维度为1的数组。

一共有2层神经网络（其中第一层必须加上“return_sequences=True”,这样才能转化成步长为3的输入变量）

（1）return_sequences=True

from keras.models import Model
from keras.layers import Input
from keras.layers import LSTM
from numpy import array
from keras.models import Sequential


data = array([0.1,0.2,0.3]).reshape((1,3,1))
inputs1 = Input(shape=(3,1))
lstm1,state_h,state_c = LSTM(2,return_sequences=True,return_state=True)(inputs1)
lstm2 = LSTM(2,return_sequences=True)(lstm1)
model = Model(input = inputs1,outputs = [lstm2])

print(model.predict(data))

输出结果为：（最后一层lstm2的每一个时间步长hidden_state的结果）

[[[0.00120299 0.0009285 ]
  [0.0040868  0.00327   ]
  [0.00869473 0.00720878]]]

(2)return_sequence = False , return_state = True

from keras.models import Model
from keras.layers import Input
from keras.layers import LSTM
from numpy import array
from keras.models import Sequential


data = array([0.1,0.2,0.3]).reshape((1,3,1))
inputs1 = Input(shape=(3,1))
lstm1,state_h,state_c = LSTM(2,return_sequences=True,return_state=True)(inputs1)
lstm2,state_h2,state_c2 = LSTM(2,return_state=True)(lstm1)
model = Model(input = inputs1,outputs = [lstm2,state_h2,state_c2])

print(model.predict(data))

输出为：

因为return_state=True，返回了最后一层最后一个时间步长的输出hidden_state和输入cell_state.

[array([[-0.00234587,  0.00718377]], dtype=float32), array([[-0.00234587,  0.00718377]], dtype=float32), array([[-0.00476015,  0.01406127]], dtype=float32)]

(3)return_sequence = True , return_state = True

from keras.models import Model
from keras.layers import Input
from keras.layers import LSTM
from numpy import array
from keras.models import Sequential


data = array([0.1,0.2,0.3]).reshape((1,3,1))
inputs1 = Input(shape=(3,1))
lstm1,state_h,state_c = LSTM(2,return_sequences=True,return_state=True)(inputs1)
lstm2,state_h2,state_c2 = LSTM(2,return_sequences=True,return_state=True)(lstm1)
model = Model(input = inputs1,outputs = [lstm2,state_h2,state_c2])

print(model.predict(data))

输出为：最后一层所有时间步长的hidden state,及最后一层最后一步的hidden state,cell state.

[array([[[-2.0248523e-04, -1.0290105e-03],
        [-3.6455912e-04, -3.3424206e-03],
        [-3.6696041e-05, -6.6624139e-03]]], dtype=float32),

 array([[-3.669604e-05, -6.662414e-03]], dtype=float32),

 array([[-7.3107367e-05, -1.3788906e-02]], dtype=float32)]

(4)return_sequence = False , return_state = False

from keras.models import Model
from keras.layers import Input
from keras.layers import LSTM
from numpy import array
from keras.models import Sequential


data = array([0.1,0.2,0.3]).reshape((1,3,1))
inputs1 = Input(shape=(3,1))
lstm1,state_h,state_c = LSTM(2,return_sequences=True,return_state=True)(inputs1)
lstm2 = LSTM(2)(lstm1)
model = Model(input = inputs1,outputs = [lstm2])

print(model.predict(data))

输出为：最后一层的最后一个步长的hidden state.

[[-0.01998264 -0.00451741]]