Python LSTM、GRU和单词嵌入问题

我一直在努力使我的模型工作，但已经三天了，我不能得到它。我正在从一个名为“体提取”的句子中提取单词 问题如下： 我有（3044，80145）矩阵作为输入，其中 3044-句子数。 80-句子中的最大字数 145-单词嵌入的尺寸（包括POS标签等）

和（3044，80）作为输出，其中： 80-一个热向量，其中1出现在相位的位置

我尝试过LSTM和GRU，但每次都（几乎）报告以下错误。请导游

TypeError: ('The following error happened while compiling the node', forall_inplace,cpu,scan_fn}(TensorConstant{80}, InplaceDimShuffle{1,0,2}.0, IncSubtensor{InplaceSet;:int64:}.0, TensorConstant{80}, gru_1_U_z, gru_1_U_r, gru_1_U_h), '\n', "Inconsistency in the inner graph of scan 'scan_fn' : an input and an output are associated with the same recurrent state and should have the same type but have type 'TensorType(float32, col)' and 'TensorType(float32, matrix)' respectively.")

我的代码是：

train_inp = pickle.load(open("train_inp_145.pkl", "rb"))
train_out = pickle.load(open("train_out_145.pkl", "rb"))

train_inp = train_inp.reshape(3044, 80, 145).astype('float32')
train_out = train_out.reshape(3044, 80, 1)

model = Sequential()

model.add(GRU(1, return_sequences=True, input_shape=(80, 145)))
model.add(TimeDistributedDense(1))
model.add(Activation("softmax"))

model.compile(loss='mse',
              optimizer='rmsprop',
              metrics=['accuracy'])

print (model.summary())
model.fit(train_inp, train_out, validation_split=0.2,
          batch_size=100,
          nb_epoch=2
          )

模型摘要：

Layer (type)                     Output Shape          Param #     Connected to                     
====================================================================================================
gru_1 (GRU)                      (None, 80, 1)         441         gru_input_1[0][0]                
____________________________________________________________________________________________________
timedistributeddense_1 (TimeDistr(None, 80, 1)         2           gru_1[0][0]                      
____________________________________________________________________________________________________
activation_1 (Activation)        (None, 80, 1)         0           timedistributeddense_1[0][0]     
====================================================================================================
Total params: 443