Python Tensorflow多对多RNN（LSTM）输出形状

我有一些序列数据（大约200k个样本，每个样本有26个时间步，有三个不同的特征）。我试着运行一个LSTM训练来预测第27个时间步中的特征

到目前为止，我的模型看起来像：

input_dx1 = Input(shape=( n_steps, 1), name='input_dx1')
input_dx2 = Input(shape=( n_steps,1), name='input_dx2')
input_dx3 = Input(shape=( n_steps,1), name='input_dx3')

lstm_dx1 = LSTM(32)(input_dx1)
lstm_dx2 = LSTM(32)(input_dx2)
lstm_dx3 = LSTM(32)(input_dx3)

output_dx1 = Dense(26, activation = 'softmax' )(lstm_dx1)
output_dx2 = Dense(26, activation = 'softmax')(lstm_dx2)
output_dx3 = Dense(26, activation = 'softmax')(lstm_dx3)

model = Model( [ input_dx1, input_dx2, input_dx3], [output_dx1, output_dx2, output_dx3])

Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_dx1 (InputLayer)          [(None, 26, 1)]      0                                            
__________________________________________________________________________________________________
input_dx2 (InputLayer)          [(None, 26, 1)]      0                                            
__________________________________________________________________________________________________
input_dx3 (InputLayer)          [(None, 26, 1)]      0                                            
__________________________________________________________________________________________________
lstm_44 (LSTM)                  (None, 32)           4352        input_dx1[0][0]                  
__________________________________________________________________________________________________
lstm_45 (LSTM)                  (None, 32)           4352        input_dx2[0][0]                  
__________________________________________________________________________________________________
lstm_46 (LSTM)                  (None, 32)           4352        input_dx3[0][0]                  
__________________________________________________________________________________________________
dense_34 (Dense)                (None, 26)           858         lstm_44[0][0]                    
__________________________________________________________________________________________________
dense_35 (Dense)                (None, 26)           858         lstm_45[0][0]                    
__________________________________________________________________________________________________
dense_36 (Dense)                (None, 26)           858         lstm_46[0][0]                    
==================================================================================================
Total params: 15,630
Trainable params: 15,630
Non-trainable params: 0
________________________________________________________________________________________________

模型摘要如下所示：

input_dx1 = Input(shape=( n_steps, 1), name='input_dx1')
input_dx2 = Input(shape=( n_steps,1), name='input_dx2')
input_dx3 = Input(shape=( n_steps,1), name='input_dx3')

lstm_dx1 = LSTM(32)(input_dx1)
lstm_dx2 = LSTM(32)(input_dx2)
lstm_dx3 = LSTM(32)(input_dx3)

output_dx1 = Dense(26, activation = 'softmax' )(lstm_dx1)
output_dx2 = Dense(26, activation = 'softmax')(lstm_dx2)
output_dx3 = Dense(26, activation = 'softmax')(lstm_dx3)

model = Model( [ input_dx1, input_dx2, input_dx3], [output_dx1, output_dx2, output_dx3])

Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_dx1 (InputLayer)          [(None, 26, 1)]      0                                            
__________________________________________________________________________________________________
input_dx2 (InputLayer)          [(None, 26, 1)]      0                                            
__________________________________________________________________________________________________
input_dx3 (InputLayer)          [(None, 26, 1)]      0                                            
__________________________________________________________________________________________________
lstm_44 (LSTM)                  (None, 32)           4352        input_dx1[0][0]                  
__________________________________________________________________________________________________
lstm_45 (LSTM)                  (None, 32)           4352        input_dx2[0][0]                  
__________________________________________________________________________________________________
lstm_46 (LSTM)                  (None, 32)           4352        input_dx3[0][0]                  
__________________________________________________________________________________________________
dense_34 (Dense)                (None, 26)           858         lstm_44[0][0]                    
__________________________________________________________________________________________________
dense_35 (Dense)                (None, 26)           858         lstm_45[0][0]                    
__________________________________________________________________________________________________
dense_36 (Dense)                (None, 26)           858         lstm_46[0][0]                    
==================================================================================================
Total params: 15,630
Trainable params: 15,630
Non-trainable params: 0
________________________________________________________________________________________________

但是当我跑的时候

from tensorflow.keras.optimizers import SGD, Adam
model.compile(optimizer=Adam(lr=0.1), loss='sparse_categorical_crossentropy',metrics='accuracy' )
model.fit([ X_tr_dx1, X_tr_dx2, X_tr_dx3], [Y_tr_dx1, Y_tr_dx2, Y_tr_dx3], epochs=1, batch_size =26)

我得到了一个错误

ValueError:cannot squeeze dim[1]，预期维度为1，对于输入形状为[？，26]的“{node squeeze}}}=squeeze[T=DT_FLOAT，squeeze_dims=[-1]]（remove_squeezable_dimensions/squeeze）得到了26

我的
Y_tr_dx1.shape
返回
（238219，26,1）
，这很有意义：我有大约200k个样本，每个样本有26个Y测量值，范围从
t2，t3…t27
到表示我在
t1，t2，…t26时提供的26个X值之后的序列中的“下一个”项。每个值本身都是一个整数编码的值，所以我使用的是
sparse\u categorical\u crossentropy

为什么我的
model.fit（）
仍然给我一个尺寸错误