python - LSTM keras - 值错误如何解析输入维度

Question

我制作了一个用于训练数据的函数，该函数选择随机文件，无论是正数还是负数。这是一个二元分类问题。

    model=Sequential()
InputBatch = np.expand_dims(InputBatch, 0)
print(InputBatch.shape)
model.add(LSTM(100,input_shape=(1,6,30),return_sequences=True))
model.compile(loss='mean_absolute_error',optimizer='adam',metrics=['accuracy'])
model.fit(InputBatch,PositiveOrNegativeLabel,batch_size=6,nb_epoch=10,verbose=1,validation_split=0.05)

InputBatch变量的形状为(1,6,30)

例如我的输入数据是

[[   nan  1520.  1295.    nan  8396.  9322. 12715.    nan  5172.  7232.
  11266.    nan 11266.  2757.  4416. 12020. 12111.     0.     0.     0.
      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]
 [   nan  3045. 11480.   900.  5842. 11496.  4463.    nan 11956.   900.
  10400.  8022.  2504. 12106.     0.     0.     0.     0.     0.     0.
      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]
 [   nan  9307. 12003.  2879.  6398.  9372.  4614.  5222.    nan    nan
   2879. 10364.  6923.  4709.  4860. 11871.     0.     0.     0.     0.
      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]
 [   nan  6689.  2818. 12003.  6480.    nan     0.     0.     0.     0.
      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.
      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]
 [   nan  3395.  1087. 11904.  7232.  8840. 10115.  4494. 11516.  7441.
   8535. 12106.     0.     0.     0.     0.     0.     0.     0.     0.
      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]
 [   nan  1287.   420.  4070. 11087.  7410. 12186.  2387. 12111.     0.
      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.
      0.     0.     0.     0.     0.     0.     0.     0.     0.     0.]]

我设置了数据的形状为(6,30)

我收到值错误

ValueError: Error when checking input: expected lstm_16_input to have 3 dimensions, but got array with shape (1,6, 30)

它正在获取三维输入，我不明白如何以及为什么

Answer 1

LSTM 输入必须位于 3 dimensional(samples,timesteps,features) 中。你的数据似乎是二维的。您可以使用 numpys reshape()将数据转换为 3D 的函数。

例如，如果您使用 1 个时间步长，则必须将其 reshape 为 array.reshape(6,1,30)或者如果您使用 6 个时间步长，则 array.reshape(1,6,30)

有关 reshape LSTM 输入的更多信息，您可以查看此 site

[[更新]] 你的代码有很多问题

 model=Sequential()
InputBatch = np.expand_dims(InputBatch, 0)
print(InputBatch.shape)
model.add(LSTM(100,input_shape=(1,6,30),return_sequences=True))
model.compile(loss='mean_absolute_error',optimizer='adam',metrics=['accuracy'])
model.fit(InputBatch,PositiveOrNegativeLabel,batch_size=6,nb_epoch=10,verbose=1,validation_split=0.05)

当你将数据转换为(1,6,30)时，你基本上是说你只有一个样本(只有1)，batch_size是6，但你只有1个样本，你只有一个样本，但你正在做验证分割。由于你只有一个 X 值，因此它也只有一个 Y(正值或负值标签)，所以我只需要一个值，即 1。

我已经运行了您的程序，并对您在问题中显示的代码和数据进行了一些更改(我将 NA 更改为 0):

    a=np.array([ 
             [0,1520,1295,0,8396,9322,12715,0,5172,7232,11266,0,11266,2757,4416,12020,12111,0,0,0,0,0,0,0,0,0,0,0,0,0],
             [0,3045,11480,900,5842,11496,4463,0,11956,900,10400,8022,2504,12106,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
             [0,9307,12003,2879,6398,9372,4614,5222,0,0,2879,10364,6923,4709,4860,11871,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
             [0,6689,2818,12003,6480,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
             [0,3395,1087,11904,7232,8840,10115,4494,11516,7441,8535,12106,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
             [0,1287,420,4070,11087,7410,12186,2387,12111,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
           ]
          )

PositiveOrNegativeLabel=np.array([[1]])
PositiveOrNegativeLabel=PositiveOrNegativeLabel.reshape(1,-1)
PositiveOrNegativeLabel.shape
InputBatch =InputBatch.reshape(1,6,30)
InputBatch.shape
model=Sequential()
model.add(LSTM(1,input_shape=(6,30)))
model.add(Dense(1,activation="sigmoid"))
model.compile(loss='mean_absolute_error',optimizer='adam',metrics=['accuracy'])
model.fit(InputBatch,PositiveOrNegativeLabel,batch_size=1,verbose=1)