Are zero padding values completely eliminated for downstream layers in LSTM?

Keras
,
#1

I’m curious about the zero-padding effect after the LSTM. I observed that when we have sequences that are 0 padded and then masked, the LSTM can handle them.

Thus, my question is whether, if I use another method as a downstream layer after LSTM, such as self-attention, Zero-padding will have no effect on self-attention and attention weights and thus will have no effect on the model’s performance?

Example of Input feeding into model:

[ 2 3.2 0.2 4.2 1.7
  3 1.0 3.2 1.0 2.3
  1 0.3 3.1 0.4 3.8
  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]

Here is the code:

def LSTM_model(input_shape, total_classes):
    input_layer = tf.keras.Input(shape=input_shape, name="time_series_activity")
    input_mask = tf.keras.layers.Masking(mask_value=0.00000)(input_layer)
      lstm_l5 = tf.keras.layers.LSTM(128, activation='tanh',
                                   recurrent_initializer=tf.keras.initializers.Orthogonal(), dropout=0.5,
                                   recurrent_dropout=0.5, return_sequences=True)(input_mask)
    lstm_l6 = tf.keras.layers.LSTM(128, activation='tanh',
                                   recurrent_initializer=tf.keras.initializers.Orthogonal(), dropout=0.9,
                                   recurrent_dropout=0.5)(lstm_l5)
    output_layer = tf.keras.layers.Dense(total_classes, activation="softmax")(lstm_l6)
    return tf.keras.models.Model(inputs=input_layer, outputs=output_layer)

train_generator = Generator(train_x, train_y, shuffle_input_segments=True)
test_generator = Generator(test_x, test_y)

input_shape = (None, train_x[0].shape[1])
model = LSTM_model(input_shape, total_classes)
model.summary()
opt = tf.keras.optimizers.Adam(learning_rate=0.001)
model.compile(loss="categorical_crossentropy",
               optimizer=opt,
               metrics=[tfa.metrics.F1Score(num_classes=total_classes, average='macro')]
               )

path_checkpoint = os.path.join(filepath,
                                dataset,
                                "best_model_checkpoint"
                                   )
callback_model_saving = tf.keras.callbacks.ModelCheckpoint(filepath=path_checkpoint,
                                                               monitor='val_f1_score',
                                                               mode='max',
                                                               verbose=1,
                                                               save_best_only=True)
history = model.fit_generator(train_generator, epochs=total_epoch, steps_per_epoch=len(train_generator),
                                  callbacks=[callbacks, callback_model_saving],
                                  validation_data=test_generator, validation_steps=len(test_generator))
#3

Hi @Nafees

No, zero padding values are not completely eliminated for downstream layers in LSTM.

LSTM layer itself can handle zero-padding through masking. Masking sets the attention weights for padded elements to zero by effectively ignoring them during the calculation of the hidden state.

Downstream layers like self-attention can still be affected by zero-padding. These layers consider all elements in the sequence when calculating their output, including the padded ones which lead to the model paying attention to irrelevant information and can potentially impacting performance.

You can use few of the ways to avoid the impact of zero-padding on downstream layers like Masking the output of LSTM layer before feeding it to downstream layers, Positional encoding, Adaptive padding and Ignoring the padded elements in the model architecture.