Keras的RNN初探

先尝试了Embedding的用法:

介绍看这里：https://keras-cn.readthedocs.io/en/latest/layers/embedding_layer/#embedding_1

from keras import backend as K
from keras.layers import Dense, Embedding, Activation, Permute,Input, Flatten, Dropout
from keras.layers.core import Lambda

def GetModel():
    inputs = Input(shape=(66,))
    net = Embedding(50,13)(inputs)
    net = Lambda(lambda x :K.expand_dims(x,-1))(net)
    net = Conv2D(7,3)(net)
    model = Model(inputs=inputs, outputs=net)
    model.summary()



'''
Layer (type)                 Output Shape              Param #   
=================================================================
input_1 (InputLayer)         (None, 66)                0         
_________________________________________________________________
embedding_1 (Embedding)      (None, 66, 13)            650       
_________________________________________________________________
lambda_1 (Lambda)            (None, 66, 13, 1)         0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 64, 11, 7)         70        
=================================================================
Total params: 720
Trainable params: 720
Non-trainable params: 0
_________________________________________________________________
'''

看了下两个包装器Wrapper,很有用:https://keras-cn.readthedocs.io/en/latest/layers/wrapper/#bidirectional

关于Keras LSTM一些参数的解释:https://blog.csdn.net/u011327333/article/details/78501054

一些试验：

def GetModel():
    inputs = Input(shape=(66,))
    net = Embedding(50,13)(inputs)
    net = Lambda(lambda x :K.expand_dims(x,-1))(net)
    net = Conv2D(7,3)(net)
    net = Reshape( (-1,77) )(net)
    net = LSTM(666)(net)
    model = Model(inputs=inputs, outputs=net)
    model.summary()

'''
=================================================================
input_1 (InputLayer)         (None, 66)                0         
_________________________________________________________________
embedding_1 (Embedding)      (None, 66, 13)            650       
_________________________________________________________________
lambda_1 (Lambda)            (None, 66, 13, 1)         0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 64, 11, 7)         70        
_________________________________________________________________
reshape_1 (Reshape)          (None, 64, 77)            0         
_________________________________________________________________
lstm_1 (LSTM)                (None, 666)               1982016   
=================================================================
'''




def GetModel():
    inputs = Input(shape=(66,))
    net = Embedding(50,13)(inputs)
    net = Lambda(lambda x :K.expand_dims(x,-1))(net)
    net = Conv2D(7,3)(net)
    net = Reshape( (-1,77) )(net)
    net = LSTM(666,return_sequences=True)(net)
    model = Model(inputs=inputs, outputs=net)
    model.summary()
'''
=================================================================
input_1 (InputLayer)         (None, 66)                0         
_________________________________________________________________
embedding_1 (Embedding)      (None, 66, 13)            650       
_________________________________________________________________
lambda_1 (Lambda)            (None, 66, 13, 1)         0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 64, 11, 7)         70        
_________________________________________________________________
reshape_1 (Reshape)          (None, 64, 77)            0         
_________________________________________________________________
lstm_1 (LSTM)                (None, 64, 666)           1982016   
=================================================================
'''






def GetModel():
    inputs = Input(shape=(66,))
    net = Embedding(50,13)(inputs)
    net = Lambda(lambda x :K.expand_dims(x,-1))(net)
    net = Conv2D(7,3)(net)
    net = Reshape( (-1,77) )(net)
    net = Bidirectional( LSTM(666,return_sequences=True,return_state=False) )(net)
    net = Bidirectional( LSTM(520,return_sequences=True,return_state=False) )(net)
    net = Dense(20)(net)
    net = Flatten()(net)

    model = Model(inputs=inputs, outputs=net)
    model.summary()
    1/0

'''
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_1 (InputLayer)         (None, 66)                0         
_________________________________________________________________
embedding_1 (Embedding)      (None, 66, 13)            650       
_________________________________________________________________
lambda_1 (Lambda)            (None, 66, 13, 1)         0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 64, 11, 7)         70        
_________________________________________________________________
reshape_1 (Reshape)          (None, 64, 77)            0         
_________________________________________________________________
bidirectional_1 (Bidirection (None, 64, 1332)          3964032   
_________________________________________________________________
bidirectional_2 (Bidirection (None, 64, 1040)          7708480   
_________________________________________________________________
dense_1 (Dense)              (None, 64, 20)            20820     
_________________________________________________________________
flatten_1 (Flatten)          (None, 1280)              0         
=================================================================
'''