Build autoencoder for single matrix with integer numbers

Question

Can you please tell me how to build an autoencoder with a single matrix(4,4) with integer numbers? I want to build an autoencoder for the below-mentioned data. I don't know whether I should convert the decimal numbers to binary first using one-hot encoding or a neural network will recognize integer numbers. e.g,

input data = array([[ 4,  3,  8,  6],
                    [ 1,  1,  2,  2],
                    [24, 18, 32, 24],
                    [ 6,  6,  8,  8]])
autoencoder(data)
output data= array([[ 4,  3,  8,  6],
                    [ 1,  1,  2,  2],
                    [24, 18, 32, 24]
                    [ 6,  6,  8,  8]])

Answer 1

You can convert your data to float32 and feed it to an stack of LSTM Autoencoder with linear activation function.

The following architecture works, but you an fine tune it:

input_data = np.array([[[4, 3, 8, 6], [1, 1, 2, 2], [24, 18, 32, 24], [6, 6, 8, 8]]]).astype(np.float32)

model = tf.keras.Sequential([
    tf.keras.layers.LSTM(4, return_sequences=True, activation="linear"),
    tf.keras.layers.LSTM(2, return_sequences=True, activation="linear"),
    tf.keras.layers.LSTM(1, return_sequences=True, activation="linear"),
    tf.keras.layers.LSTM(2, return_sequences=True, activation="linear"),
    tf.keras.layers.LSTM(4, return_sequences=True, activation="linear"),
])

model.compile(optimizer="Adam", loss="MAE", metrics=["MSE"])
model.fit(input_data, input_data, epochs=200000)

What I have got after 200,000 epochs!:

[[[ 3.9911306   2.9956746   7.989547    5.9939675 ]
  [ 0.9976706   0.99182177  2.0006196   1.9999651 ]
  [23.9897     17.996624   31.995388   24.006733  ]
  [ 6.0549197   5.9584255   8.023056    8.003849  ]]]