keras张量整形（lstm输入形状错误）

Question

I am using LSTM on keras and using a reshape layer prior in hopes that I don't have to specify shape for the LSTM layer.

the input is 84600 x 6

84600 seconds in 2 months. 6 metric/[labels] im measuring throughout the 2 months

so far I have

model = tf.keras.models.Sequential()
model.add(tf.keras.layers.Reshape((86400,1,6), input_shape=(84600, 6)))
model.add(tf.keras.layers.LSTM(128,  activation='relu', input_shape= 
(x_train.shape), return_sequences=True))
model.add(tf.keras.layers.Dense(10, activation='softmax'))

which throws an error:

ValueError: Input 0 of layer lstm is incompatible with the layer: expected ndim=3, found ndim=4. Full shape received: [None, 86400, 1, 6]

this is understandable. The batch size, plus 3 layers equals 4. However, when I reshape

model.add(tf.keras.layers.Reshape((86400,1,6), input_shape=(84600, 6)))
vvvvvvv
model.add(tf.keras.layers.Reshape((86400,6), input_shape=(84600, 6)))

It throws

ValueError: Error when checking input: expected reshape_input to have 3 dimensions, but got array with shape (86400, 6)

It seems to ignore the batch size as an array element. And treats it as 2 indexes. It jumps from 4 dimensions to 2 dimensions.

The problem is LSTM takes 3 dimensions as input, and I can't seem to get that. Ideally I want a 86400 x 1 x 6 array/tensor. So it becomes 84600 examples of 1x6 data.

Thank you very much!

Answer 1

The problem is that the way you are reshaping your input is incompatible with an LSTM layer. An LSTM layer expects an input with 3 dimensions: (batch_size, timesteps, features) . However, you are feeding it an input with shape (batch_size, 84600, 1, 6) .

In your case it seems like 84600 is the number of timesteps and 6 is the number of features per timestep. So, it makes more sense to leave out the Reshape layer and simply use input_shape (84600, 6) for your LSTM layer:

model = tf.keras.models.Sequential()
model.add(tf.keras.layers.LSTM(128,  activation='relu', input_shape=(84600, 6), return_sequences=True))
model.add(tf.keras.layers.Dense(10, activation='softmax'))