def test_concat_sparse_dense_correct_output_for_dense_input() -> None:
layer = ConcatenateSparseDenseFeatures(
attribute=attribute_name,
feature_type=SEQUENCE,
feature_type_signature=[
realistic_feature_signature_dense_1,
realistic_feature_signature_dense_2,
],
config=dict(
model_config_basic,
# also activate all dropout to check that it has no effect on dense features
**{
SPARSE_INPUT_DROPOUT: True,
DENSE_INPUT_DROPOUT: True
},
),
)
outputs_expected = [
[[10.0, 1.0, 2.0], [20.0, 3.0, 4.0], [30.0, 5.0, 6.0]],
[[40.0, 1.5, 2.5], [50.0, 3.5, 4.5], [0.0, 0.0, 0.0]],
]
inputs = ([realistic_feature_dense_seq_1, realistic_feature_dense_seq_2], )
train_outputs = layer(inputs, training=True)
assert (train_outputs.numpy() == outputs_expected).all()
test_outputs = layer(inputs, training=False)
assert (test_outputs.numpy() == outputs_expected).all()
def test_concat_sparse_dense_applies_dropout_to_sparse_densified_input(
) -> None:
layer_dropout_for_sparse_densified = ConcatenateSparseDenseFeatures(
attribute=attribute_name,
feature_type=SEQUENCE,
feature_type_signature=[
feature_signature_sparse_1, feature_signature_sparse_1
],
config=dict(model_config_basic, **{
DENSE_INPUT_DROPOUT: True,
DROP_RATE: 0.99999999
}), # keras dropout doesn't accept velues >= 1.0
)
inputs = ([feature_sparse_seq_1, feature_sparse_seq_1], )
expected_outputs_train = tf.zeros(
(batch_size, max_seq_length, units_sparse_to_dense * 2))
train_outputs = layer_dropout_for_sparse_densified(inputs, training=True)
assert np.allclose(train_outputs.numpy(), expected_outputs_train.numpy())
# We can't check exact output contents for sparse inputs but during test-time no
# dropout should be applied, hence the outputs should not be all zeros in this case
# (unlike at training time).
test_outputs = layer_dropout_for_sparse_densified(inputs, training=False)
assert not np.allclose(test_outputs.numpy(),
expected_outputs_train.numpy())
def test_concat_sparse_dense_raises_exception_when_inconsistent_sparse_features() -> None: # noqa: E501
with pytest.raises(TFLayerConfigException):
ConcatenateSparseDenseFeatures(
attribute=attribute_name,
feature_type=SEQUENCE,
feature_type_signature=[
FeatureSignature(is_sparse=True, units=2, number_of_dimensions=3),
FeatureSignature(is_sparse=True, units=1, number_of_dimensions=3),
],
config=model_config_basic,
)