检查目标时出错：预期 dense_21 有 2 个维度，但得到了形状为 (2300, 4, 54) 的数组

Question

当给定一个预测变量向量时，我遇到了一个处理预测四个输出的问题。它在 LSTM 层输入处引发错误。

我有

X.shape,Y.shape = ((2300, 36, 768), (2300, 4, 54))


# Core Part

checkpoint = ModelCheckpoint('model-{epoch:03d}-{acc:03f}-{val_acc:03f}.h5', verbose=1, monitor='val_loss',save_best_only=True, mode='auto')  

data_dim = 768
timesteps = X.shape[1]
num_classes = 10


# # expected input data shape: (batch_size, timesteps, data_dim)
model = Sequential()

model.add(LSTM(256, return_sequences=True, input_shape=(timesteps, data_dim))) 
model.add(LSTM(64, return_sequences=True))  
model.add(LSTM(32)) 

model.add(Dense(54, activation='softmax'))

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

model.fit(X, Y, batch_size=64, epochs=10, validation_split=0.1)

当我有X.shape,Y.shape = ((2300, 36, 768), (2300, 15))时，上面的代码工作正常

我该如何克服这个问题，如果我有超过一个、四个或十个预测输出，我该如何设置 LSTM 和 DENSE 层？

提前致谢。

Answer 1

如果你的输出不是一个序列，你可以使用Reshape层如下：

from keras.models import Sequential
from keras.layers import LSTM, Reshape, Dense, Activation
model = Sequential()

model.add(LSTM(256, return_sequences=True, input_shape=(100, 32))) 
model.add(LSTM(64, return_sequences=True))  
model.add(LSTM(32))

# now we have a vector of length 32 and
# we want the dense layer to have 4*54=216 neurons so that
# we will be able to reshape it to a matrix of shape (4, 54) in
# the next (Reshape) layer.
model.add(Dense(4*54))
model.add(Reshape((4, 54)))

model.add(Activation('softmax'))

model.summary()

将打印：

Model: "sequential_1"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
lstm_1 (LSTM)                (None, 100, 256)          295936    
_________________________________________________________________
lstm_2 (LSTM)                (None, 100, 64)           82176     
_________________________________________________________________
lstm_3 (LSTM)                (None, 32)                12416     
_________________________________________________________________
dense_1 (Dense)              (None, 216)               7128      
_________________________________________________________________
reshape_1 (Reshape)          (None, 4, 54)             0         
_________________________________________________________________
activation_1 (Activation)    (None, 4, 54)             0         
=================================================================
Total params: 397,656
Trainable params: 397,656
Non-trainable params: 0

Answer 2

如果您希望获得 4 个输出，那么您的最终 Dense 层应该有 4 个神经元，而不是 54，即，

model.add(Dense(54, activation='softmax'))

应该是：

model.add(Dense(4, activation='softmax'))

另外，您的 Y 的形状应该是 (number_of_examples, 4)，在您的情况下应该是

(2300, 4) 代替 (2300, 4, 54)

如果您实际上有一个形状为(2300, 4, 54) 的Y，那么您不能在输出中使用Dense 层，因为Dense 层仅适用于一维示例每个向量。