在张量流中使用 LSTM 层时出现 ValueError 问题

Question

I am trying to build up a very simple LSTM structure using padding and masking to learn how to train time series data. Suppose I have two people's blood value information per month. For the first person, I only have seven months of data, wheares, for the next person, I only have four months of data. Lastly, I select -100 as my padding value to complete the sequences.

import numpy as np
import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense
from tensorflow.keras.layers import LSTM
from tensorflow.keras.utils import Sequence
from keras.layers import LSTM, Dense, Masking

#each column represent a different blood value. 
xPad =  np.array([[
  [  0.4654949,    0.06225133],
  [ -0.48630088,   0.97063685],
  [ -0.23714237,  1.07598604],
  [ -0.94519772,   0.76515959],
  [ -0.81456729,   1.05963647],
  [  0.60236851,   1.26799774],
  [  1.89095161,   1.02534057]],

 [[ -1.76505643,   0.61171791],
  [  2.00335928,  -0.02941931],
  [ -1.58293956,  -0.02671103],
  [  1.57166957,  -0.39450184],
  [-100,         -100        ],
  [-100,         -100        ],
  [-100,         -100        ]]])

Suppose I would like to estimate whether the person will be low weight, medium weight, and high weight. That's why I also have this information for each month per person as shown below. For the second person, I fill the missing months with zero vector. For instance, [0,0,1] implies that the person within that spesific month is high weight.

y = np.array([
    [[0,0,1],[0,0,1],[0,0,1],[0,0,1],[0,1,0],[0,1,0],[0,1,0]],
    [[1,0,0],[1,0,0],[1,0,0],[0,1,0],[0,0,0],[0,0,0],[0,0,0]]
    ])

Then, I create my network using tensorflow.

special_value= -100
seq_len = 7
dim = 2

model = Sequential()
model.add(Masking(mask_value=special_value, input_shape=(seq_len, dim)))
model.add(LSTM(5))
model.add(Dense(3))
model.compile(loss='binary_crossentropy', optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001))
model.fit(xPad, y, epochs=10, batch_size=2)

I was wondering what I am doing wrong here. I would appreciate any help.

Dimensions must be equal, but are 7 and 2 for '{{node binary_crossentropy/mul}} = Mul[T=DT_FLOAT](binary_crossentropy/Cast, binary_crossentropy/Log)' with input shapes: [2,7,3], [2,3]

Answer 1

The short answer is you can make your problem go away by changing

model.add(LSTM(5))

to,

model.add(LSTM(5, return_sequences=True))

Here's an expansion on the answer.

You got xPad of shape [2, 7, 2] and y of shape [2, 7, 3] . So it seems you are trying to predict y for every time step of xPad . Then what the error says is,

It has an LSTM that outputs only the last state of it, but your y is for all of the time steps (which are incompatible). return_sequences=True outputs the states for all of the time steps, which makes your y compatible with the output of the model.

This answer shows what return_sequences=True does.