当返回顺序为真时，Keras尺寸必须相等

Question

我试图建立一个模型，根据一系列的3个观察结果预测4个值，即：

如果以下是数据

+--------------------------------+
|feature |feature |feature |Value|
+--------------------------------+
|0.1     |0.1     |0.1     |1    |
+--------------------------------+
|0.2     |0.2     |0.2     |2    |
+--------------------------------+
|0.3     |0.3     |0.3     |3    |
+--------------------------------+
|0.4     |0.4     |0.4     |4    |
+--------+--------+--------+-----+

我想根据

+--------------------------+
|feature |feature |feature |
+--------------------------+
|0.1     |0.1     |0.1     |
+--------------------------+
|0.2     |0.2     |0.2     |
+--------------------------+
|0.3     |0.3     |0.3     |
+--------+--------+--------+

我的X,y形状如下(1228, 3, 19) (1228, 4, 1)

---

def get_model():
  model = Sequential()
  model.add(LSTM(32, activation='tanh', return_sequences=True, input_shape=(X.shape[1], X.shape[2]))),
  model.add(Dense(32, activation='relu')),
  model.add(Dense(4, activation='sigmoid'))

  model.compile(loss='mse', optimizer="adam", metrics=['mae', 'mse'])
  return model

---

My Model code:
Model: "sequential_17"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
lstm_17 (LSTM)               (None, 3, 32)             6656      
_________________________________________________________________
dense_34 (Dense)             (None, 3, 32)             1056      
_________________________________________________________________
dense_35 (Dense)             (None, 3, 4)              132       
=================================================================
Total params: 7,844
Trainable params: 7,844
Non-trainable params: 0
_________________________________________________________________

当我尝试拟合数据时：

history = model.fit(X_train, y_train, epochs=200, batch_size=64, validation_split=0.2, verbose=0, callbacks=[tensorboard_callback])

我得到以下错误：

ValueError: Dimensions must be equal, but are 3 and 4 for

我应该如何重塑数据以使其正常工作，我应该填充缺失的序列吗

Answer 1

如果理解正确，每个示例都有以下内容：

输入->；(3,19) 产出->；（4,1）

其中，您尝试基于3个19个值的序列回归4个值。如果这是正确的，那么您可以在模型中使用return_sequences=False，并将输出（y）重塑为（4，）的形状，而不是像y=np.squeeze(y, -1)那样的（4,1）。或者，如果要保留序列，请使用TimeDistributed和GlobalAveragePooling1D层，并对输出执行相同的操作。它看起来是这样的：

model = Sequential()
model.add(LSTM(32, activation='tanh', return_sequences=True, input_shape=(3, 19))),
model.add(TimeDistributed(Dense(32, activation='relu')))
model.add(GlobalAveragePooling1D())
model.add(Dense(4, activation='sigmoid'))

model.summary()

Model: "sequential_9"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
lstm_9 (LSTM)                (None, 3, 32)             6656      
_________________________________________________________________
time_distributed_4 (TimeDist (None, 3, 32)             1056      
_________________________________________________________________
global_average_pooling1d_2 ( (None, 32)                0         
_________________________________________________________________
dense_16 (Dense)             (None, 4)                 132       
=================================================================
Total params: 7,844
Trainable params: 7,844
Non-trainable params: 0

编辑

当前模型的问题在于，它期望目标/输出的形状为(3,4)，而实际输出的形状为(4,1)