Keras masking example

mask_example.py

from keras.layers import Masking, Dense
from keras.layers.recurrent import LSTM
from keras.models import Sequential

import numpy as np

np.set_printoptions(precision=4)
np.random.seed(1)


def print_output_val(out_val):
    print("\n'--> time'")
    np.linspace(0, n_timesteps - 1, n_timesteps)
    for sample in range(n_samples):
        print('# sample = ', sample)
        print('input:')
        print(data[sample, :, :].T)
        print('output_val:')
        print(out_val[sample, :, :].T)


def sequential_non_temporal_example():
    model = Sequential()
    model.add(Masking(mask_value=mask_value, input_shape=(n_timesteps, n_features)))
    model.add(Dense(1, activation='linear', kernel_initializer="one"))

    output_val = model.predict(data)
    print_output_val(output_val)


def sequential_temporal_example():
    model = Sequential()
    model.add(Masking(mask_value=mask_value, input_shape=(n_timesteps, n_features)))
    model.add(LSTM(2, return_sequences=True, kernel_initializer="one"))
    model.add(Dense(1, activation='linear', kernel_initializer="one"))

    output_val = model.predict(data)
    print_output_val(output_val)


n_samples = 3
n_timesteps = 7
n_features = 2
mask_value = np.NaN  # -999999999.0# -1.0 # -1 # 0.0
data = np.ones((n_samples, n_timesteps, n_features))

for s in range(n_samples):
    for f in range(n_features):
        data[s, :, f] = np.linspace(1, n_timesteps, n_timesteps)

# mask a feature value of one sample and timestep (no effect)
data[1, 0, 0] = mask_value

# mask all feature values of one sample and timestep (propagates 0.*mask_value at layer of step/sample?)
data[2, 3, :] = mask_value

print('####################### sequential_non_temporal_example #######################:')
sequential_non_temporal_example()
print('####################### sequential_temporal_example #######################:')
# As non-temporal but masked timestep state does not propagate through time:
sequential_temporal_example()

# ####################### sequential_non_temporal_example #######################:
# _________________________________________________________________
# Layer (type)                 Output Shape              Param #
# =================================================================
# masking_1 (Masking)          (None, 7, 2)              0
# _________________________________________________________________
# dense_1 (Dense)              (None, 7, 1)              3
# =================================================================
# Total params: 3.0
# Trainable params: 3.0
# Non-trainable params: 0.0
# _________________________________________________________________
# --> time
# [ 0.  1.  2.  3.  4.  5.  6.]
# # sample =  0
# input:
# [[ 1.  2.  3.  4.  5.  6.  7.]
#  [ 1.  2.  3.  4.  5.  6.  7.]]
# output_val:
# [[  2.   4.   6.   8.  10.  12.  14.]]
# # sample =  1
# input:
# [[ nan   2.   3.   4.   5.   6.   7.]
#  [  1.   2.   3.   4.   5.   6.   7.]]
# output_val:
# [[ nan   4.   6.   8.  10.  12.  14.]]
# # sample =  2
# input:
# [[  1.   2.   3.  nan   5.   6.   7.]
#  [  1.   2.   3.  nan   5.   6.   7.]]
# output_val:
# [[  2.   4.   6.  nan  10.  12.  14.]]
# ####################### sequential_temporal_example #######################:
# _________________________________________________________________
# Layer (type)                 Output Shape              Param #
# =================================================================
# masking_2 (Masking)          (None, 7, 2)              0
# _________________________________________________________________
# lstm_1 (LSTM)                (None, 7, 2)              40
# _________________________________________________________________
# dense_2 (Dense)              (None, 7, 1)              3
# =================================================================
# Total params: 43.0
# Trainable params: 43.0
# Non-trainable params: 0.0
# _________________________________________________________________
# --> time
# [ 0.  1.  2.  3.  4.  5.  6.]
# # sample =  0
# input:
# [[ 1.  2.  3.  4.  5.  6.  7.]
#  [ 1.  2.  3.  4.  5.  6.  7.]]
# output_val:
# [[ 1.2603  1.9066  1.9871  1.9982  1.9998  2.      2.    ]]
# # sample =  1
# input:
# [[ nan   2.   3.   4.   5.   6.   7.]
#  [  1.   2.   3.   4.   5.   6.   7.]]
# output_val:
# [[ nan  nan  nan  nan  nan  nan  nan]]
# # sample =  2
# input:
# [[  1.   2.   3.  nan   5.   6.   7.]
#  [  1.   2.   3.  nan   5.   6.   7.]]
# output_val:
# [[ 1.2603  1.9066  1.9871     nan     nan     nan     nan]]