如何在Keras中使用张量的一部分？

Question

I want to use part of a the tensor in the output of lstm layer, but don't know how to do it correctly. My purpose is tell LSTM layer the "real" length of its input sequence. Here is my attempt, but it fails. Isthere anyone who can help solve this problem and explain the details, thanks a lot~

input_spectrogram = Input(shape=(64,500,1))
input_length = Input(shape=(1,))
cnn1 = Conv2D(filters = 64, kernel_size = (1,4),input_shape=(64,500, 1),padding = 'same',strides = 1,activation = 'relu',name='conv1')(input_spectrogram)
maxpooling1 = MaxPooling2D(pool_size = (1,4),name='maxpooling1')(cnn1)
bn1 = BatchNormalization(name='BN1')(maxpooling1)
cnn2 = Conv2D(filters = 128, kernel_size = (64,1),strides = 1,activation ='relu',name='conv2')(bn1)
maxpooling2 = MaxPooling2D(pool_size = (1,2),name='maxpooling2')(cnn2)
reshape = Reshape((62,128))(maxpooling2)
lstm1 = LSTM(128,return_sequences = True,recurrent_dropout=0.3,name='lstm1')(reshape)   #output:(None,62,128)
softmax_in = Lambda(lambda x:x[0][x[1],:])([lstm1,input_length])
softmax_ = Dense(10,activation='softmax',name='softmax_')(softmax_in)
seq = Model(inputs=input_spectrogram, outputs=[softmax_])
seq.compile(loss='categorical_crossentropy', optimizer='adadelta',metrics=['accuracy'])

Answer 1

Seems to be indexing with tensor is not fully supported (see discussion here: https://github.com/tensorflow/tensorflow/issues/206#issuecomment-158435464 ).

Does it work for you to perform indexing with constant instead?

Answer 2

input_spectrogram = Input(shape=(64,500,1))
input_length = Input(shape=(1,))
cnn1 = Conv2D(filters = 64, kernel_size = (1,4),input_shape=(64,500, 1),padding = 'same',strides = 1,activation = 'relu',name='conv1')(input_spectrogram)
maxpooling1 = MaxPooling2D(pool_size = (1,4),name='maxpooling1')(cnn1)
bn1 = BatchNormalization(name='BN1')(maxpooling1)
cnn2 = Conv2D(filters = 128, kernel_size = (64,1),strides = 1,activation ='relu',name='conv2')(bn1)
maxpooling2 = MaxPooling2D(pool_size = (1,2),name='maxpooling2')(cnn2)
reshape = Reshape((62,128))(maxpooling2)
lstm1 = LSTM(128,return_sequences = True,recurrent_dropout=0.3,name='lstm1')(reshape)   #output:(None,62,128)
softmax_in = Lambda(lambda x:x[:,5])(lstm1)
softmax_ = Dense(10,activation='softmax',name='softmax_')(softmax_in)
seq = Model(inputs=input_spectrogram, outputs=[softmax_])
seq.compile(loss='categorical_crossentropy', optimizer='adadelta',metrics=['accuracy'])

now it is feasible, so how to use the "real_length" from an input layer?