我正在尝试在Keras中为视频构建编码器-解码器体系结构。我有一个卷积LSTM，如果我具有相同大小的输入和输出，则编码器和解码器可以正常工作，但是，在输出大小与输入长度不同的情况下，我无法使其工作。
来自另一个stackoverflow问题的代码本身，尽管与自动编码器有关。我现在希望有一个用于目标序列的解码器。
下面的代码工作正常，但是如何更改最终输出，解码后的输出时间不长48个步长，例如说20个。
input_seq = Input(shape=(48,32,16,1))
a = ConvLSTM2D(40, (3, 3), activation='sigmoid', padding='same', return_sequences=True)(input_seq)
a = ConvLSTM2D(40, (3, 3), activation='sigmoid', padding='same', return_sequences=True )(a)
b = MaxPooling3D((2,2,2), padding='same')(a)
c = ConvLSTM2D(40, (3, 3), activation='sigmoid', padding='same', return_sequences=True)(b)
c = ConvLSTM2D(40, (3, 3), activation='sigmoid', padding='same',return_sequences=True)(c)
encoded = MaxPooling3D((2,2,2), padding='same', name="encoder")(c)
d = ConvLSTM2D(40, (3, 3), activation='relu', padding='same',return_sequences=True )(encoded)
d = ConvLSTM2D(40, (3, 3), activation='relu', padding='same', return_sequences=True)(d)
e= UpSampling3D((2, 2,2))(d)
##Skip connection
#merge_one = concatenate([b, e])
f = ConvLSTM2D(40, (3, 3), activation='sigmoid', padding='same', return_sequences=True) (e)#(e)
f = ConvLSTM2D(40, (3, 3), activation='sigmoid', padding='same', return_sequences=True) (f)#(e)
g = UpSampling3D((2, 2,2))(f)
decoded = Conv3D(1, (3, 3, 3), activation='sigmoid', padding='same')(g)# (merge_two)
model = Model(input_seq, decoded)
model.compile(optimizer='adadelta', loss='binary_crossentropy')
model.summary()
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_4 (InputLayer) (None, 48, 32, 16, 1) 0
_________________________________________________________________
conv_lst_m2d_17 (ConvLSTM2D) (None, 48, 32, 16, 40) 59200
_________________________________________________________________
conv_lst_m2d_18 (ConvLSTM2D) (None, 48, 32, 16, 40) 115360
_________________________________________________________________
max_pooling3d_3 (MaxPooling3 (None, 24, 16, 8, 40) 0
_________________________________________________________________
conv_lst_m2d_19 (ConvLSTM2D) (None, 24, 16, 8, 40) 115360
_________________________________________________________________
conv_lst_m2d_20 (ConvLSTM2D) (None, 24, 16, 8, 40) 115360
_________________________________________________________________
encoder (MaxPooling3D) (None, 12, 8, 4, 40) 0
_________________________________________________________________
conv_lst_m2d_21 (ConvLSTM2D) (None, 12, 8, 4, 40) 115360
_________________________________________________________________
conv_lst_m2d_22 (ConvLSTM2D) (None, 12, 8, 4, 40) 115360
_________________________________________________________________
up_sampling3d_5 (UpSampling3 (None, 24, 16, 8, 40) 0
_________________________________________________________________
conv_lst_m2d_23 (ConvLSTM2D) (None, 24, 16, 8, 40) 115360
_________________________________________________________________
conv_lst_m2d_24 (ConvLSTM2D) (None, 24, 16, 8, 40) 115360
_________________________________________________________________
up_sampling3d_6 (UpSampling3 (None, 48, 32, 16, 40) 0
_________________________________________________________________
conv3d_3 (Conv3D) (None, 48, 32, 16, 1) 1081
=================================================================
Total params: 867,801
Trainable params: 867,801
Non-trainable params: 0
_________________________________________________________________