Python 时间分布层：不兼容的形状：[32,3,3]与[32,3]

我有一个模型，以前是有效的，但现在它不工作。它给出了一个关于以下内容的错误：

InvalidArgumentError:不兼容的形状：[32,3,3]与[32,3] [{{node-Nadam/gradients/loss/time_-distributed_-loss/SquaredDifference_-grad/BroadcastGradientArgs}]

旧型号：

class Model:
def set_the_model(self, look_back):
    self.model = Sequential()
    self.model.add(LSTM(16, activation="relu", input_shape=(look_back, 3),return_sequences=True)) #, stateful=True
    self.model.add(Dropout(0.2))
    self.model.add(LSTM(32, activation="relu", return_sequences=True))
    self.model.add(Dropout(0.2))
    self.model.add(TimeDistributed(Dense(look_back)))
    #self.model.add(Dense(3))
    self.model.compile(loss="mse", optimizer="nadam", metrics=['acc']) #mse
    self.model.summary()

def start_train(self, trainD1, trainD2):
    es = EarlyStopping(monitor='loss', patience = 2, mode='min')
    #for i, j in trainD1, trainD2:
    self.model.fit(trainD1, trainD2, epochs=200, batch_size = 32, verbose=1,  callbacks=[es])
    #  self.model.reset_states()

def predict_result(self, test_case):
    value = self.model.predict(test_case, verbose=0)
    return value

当我在

start\u train

方法中取消注释两行代码时，错误开始出现，我将代码转换为以下内容（当前模型）：

然后，我想回到旧型号。当我开始训练模型时，它给出了上面的错误

这是模型的外观：

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
lstm (LSTM)                  (None, 3, 16)             1280      
_________________________________________________________________ 
dropout (Dropout)            (None, 3, 16)             0         
_________________________________________________________________
lstm_1 (LSTM)                (None, 3, 32)             6272      
_________________________________________________________________
dropout_1 (Dropout)          (None, 3, 32)             0         
_________________________________________________________________
time_distributed (TimeDistri (None, 3, 3)              99        
=================================================================
Total params: 7,651
Trainable params: 7,651
Non-trainable params: 0

---

编辑：在旧模型上，我将批处理大小更改为1，现在可以工作了，但现在速度会非常慢。

在数据集中循环每个观察值的目的是什么？正如我正确理解的那样，这正是I，j在zip（trainD1，trainD2）所做的。这就是为什么您的代码处理的是batch_size=1，而不处理较高的batch_size值。@s3nh“旧模型”处理的是单个数据集。第二个模型适用于多个输入，如果我将这些输入组合起来，它将不好（根据我的数据）。我处理的每个输入都不具有相同的输入大小；因此，不能除以32。如果我不作为一个发送，它会给我一个错误

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
lstm (LSTM)                  (None, 3, 16)             1280      
_________________________________________________________________ 
dropout (Dropout)            (None, 3, 16)             0         
_________________________________________________________________
lstm_1 (LSTM)                (None, 3, 32)             6272      
_________________________________________________________________
dropout_1 (Dropout)          (None, 3, 32)             0         
_________________________________________________________________
time_distributed (TimeDistri (None, 3, 3)              99        
=================================================================
Total params: 7,651
Trainable params: 7,651
Non-trainable params: 0