python - Keras lstm with masking layer for variable-length inputs

Question

I know this is a subject with a lot of questions but I couldn't find any solution to my problem.

I am training a LSTM network on variable-length inputs using a masking layer but it seems that it doesn't have any effect.

Input shape (100, 362, 24) with 362 being the maximum sequence lenght, 24 the number of features and 100 the number of samples (divided 75 train / 25 valid).

Output shape (100, 362, 1) transformed later to (100, 362 - N, 1).

Here is the code for my network:

from keras import Sequential
from keras.layers import Embedding, Masking, LSTM, Lambda
import keras.backend as K


#                          O O O
#   example for N:3        | | |
#                    O O O O O O
#                    | | | | | | 
#                    O O O O O O

N = 5
y= y[:,N:,:]

x_train = x[:75]
x_test = x[75:]
y_train = y[:75]
y_test = y[75:]

model = Sequential()
model.add(Masking(mask_value=0., input_shape=(timesteps, features)))
model.add(LSTM(128, return_sequences=True))
model.add(LSTM(64, return_sequences=True))
model.add(LSTM(1, return_sequences=True))
model.add(Lambda(lambda x: x[:, N:, :]))

model.compile('adam', 'mae')

print(model.summary())
history = model.fit(x_train, y_train, 
                    epochs=3, 
                    batch_size=15, 
                    validation_data=[x_test, y_test])

my data is padded at the end. example:

>> x_test[10,350]
array([0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
   0., 0., 0., 0., 0., 0., 0.], dtype=float32)

The problem is that the mask layer seems to have no effect. I can see it with the loss value being printed during training which is equal to the one without mask I calculate after:

Layer (type)                 Output Shape              Param #   
=================================================================
masking_1 (Masking)          (None, 362, 24)           0         
_________________________________________________________________
lstm_1 (LSTM)                (None, 362, 128)          78336     
_________________________________________________________________
lstm_2 (LSTM)                (None, 362, 64)           49408     
_________________________________________________________________
lstm_3 (LSTM)                (None, 362, 1)            264       
_________________________________________________________________
lambda_1 (Lambda)            (None, 357, 1)            0         
=================================================================
Total params: 128,008
Trainable params: 128,008
Non-trainable params: 0
_________________________________________________________________
None
Train on 75 samples, validate on 25 samples
Epoch 1/3
75/75 [==============================] - 8s 113ms/step - loss: 0.1711 - val_loss: 0.1814
Epoch 2/3
75/75 [==============================] - 5s 64ms/step - loss: 0.1591 - val_loss: 0.1307
Epoch 3/3
75/75 [==============================] - 5s 63ms/step - loss: 0.1057 - val_loss: 0.1034

>> from sklearn.metrics import mean_absolute_error
>> out = model.predict(x_test, batch_size=1)
>> print('wo mask', mean_absolute_error(y_test.ravel(), out.ravel()))
>> print('w mask', mean_absolute_error(y_test[~(x_test[:,N:] == 0).all(axis=2)].ravel(), out[~(x_test[:,N:] == 0).all(axis=2)].ravel()))
wo mask 0.10343371
w mask 0.16236152

Futhermore, if I use nan value for the masked output values, I can see the nan being propagated during training (loss equals nan).

What am I missing to make the masking layer work as expected?

Answer 1

The Lambda layer, by default, does not propagate masks. In other words, the mask tensor computed by the Masking layer is thrown away by the Lambda layer, and thus the Masking layer has no effect on the output loss.

If you want the compute_mask method of a Lambda layer to propagate previous mask, you have to provide the mask argument when the layer is created. As can be seen from the source code of Lambda layer,

def __init__(self, function, output_shape=None,
             mask=None, arguments=None, **kwargs):
    # ...
    if mask is not None:
        self.supports_masking = True
    self.mask = mask

# ...

def compute_mask(self, inputs, mask=None):
    if callable(self.mask):
        return self.mask(inputs, mask)
    return self.mask

Because the default value of mask is None, compute_mask returns None and the loss is not masked at all.

To fix the problem, since your Lambda layer itself does not introduce any additional masking, the compute_mask method should just return the mask from the previous layer (with appropriate slicing to match the output shape of the layer).

masking_func = lambda inputs, previous_mask: previous_mask[:, N:]
model = Sequential()
model.add(Masking(mask_value=0., input_shape=(timesteps, features)))
model.add(LSTM(128, return_sequences=True))
model.add(LSTM(64, return_sequences=True))
model.add(LSTM(1, return_sequences=True))
model.add(Lambda(lambda x: x[:, N:, :], mask=masking_func))

Now you should be able to see the correct loss value.

>> model.evaluate(x_test, y_test, verbose=0)
0.2660679519176483
>> out = model.predict(x_test)
>> print('wo mask', mean_absolute_error(y_test.ravel(), out.ravel()))
wo mask 0.26519736809498456
>> print('w mask', mean_absolute_error(y_test[~(x_test[:,N:] == 0).all(axis=2)].ravel(), out[~(x_test[:,N:] == 0).all(axis=2)].ravel()))
w mask 0.2660679670482195

Using NaN value for padding does not work because masking is done by multiplying the loss tensor with a binary mask (0 * nan is still nan, so the mean value would be nan).