def test_meanpool_agg_apply_no_bias():
# By default, bias_initializers="zeros", so all bias terms are initialised to zeros.
agg = MeanPoolingAggregator(2, act="linear", kernel_initializer="ones")
assert agg.get_config()["kernel_initializer"]["class_name"] == "Ones"
assert agg.get_config()["bias_initializer"]["class_name"] == "Zeros"
# Self features
inp1 = keras.Input(shape=(1, 2))
# Neighbour features
inp2 = keras.Input(shape=(1, 2, 2))
out = agg([inp1, inp2])
# Check sizes
assert agg.weight_dims == [1, 1]
# Numerical test values
x1 = np.array([[[1, 1]]])
x2 = np.array([[[[2, 2], [3, 3]]]])
# Agg output:
# neigh_agg = mean(relu(x2 · ones(2x2) + zeros(2)), axis=1)
# = mean([[4,4],[6,6]]) = [[5,5]]
# from_self = K.dot(x1, ones) = [[2]]
# from_neigh = K.dot(neigh_agg, ones) = [[10]]
model = keras.Model(inputs=[inp1, inp2], outputs=out)
actual = model.predict([x1, x2])
expected = np.array([[[2, 10]]])
assert expected == pytest.approx(actual)
def test_meanpool_agg_apply_hidden_bias():
# Specifying bias_initializer="ones" initialises all bias terms to ones;
# using bias=False turns of outer bias but retains hidden bias.
agg = MeanPoolingAggregator(
2, bias=False, act="linear", kernel_initializer="ones", bias_initializer="ones"
)
assert agg.get_config()["kernel_initializer"]["class_name"] == "Ones"
assert agg.get_config()["bias_initializer"]["class_name"] == "Ones"
# Self features
inp1 = keras.Input(shape=(1, 2))
# Neighbour features
inp2 = keras.Input(shape=(1, 2, 2))
out = agg([inp1, inp2])
# Check sizes
assert agg.weight_dims == [1, 1]
# Numerical test values
x1 = np.array([[[1, 1]]])
x2 = np.array([[[[2, 2], [3, 3]]]])
# Agg output:
# neigh_agg = mean(relu(x2 · ones(2x2) + ones(2)), axis=1)
# = mean([[5,5],[7,7]]) = [[6,6]]
# from_self = K.dot(x1, ones) = [[2]]
# from_neigh = K.dot(neigh_agg, ones(2x1)) = [[12]]
model = keras.Model(inputs=[inp1, inp2], outputs=out)
actual = model.predict([x1, x2])
expected = np.array([[[2, 12]]])
assert expected == pytest.approx(actual)
def test_meanpool_agg_constructor():
agg = MeanPoolingAggregator(2, bias=False)
assert agg.output_dim == 2
assert agg.hidden_dim == 2
assert not agg.has_bias
assert agg.act.__name__ == "relu"
assert agg.hidden_act.__name__ == "relu"
# Check config
config = agg.get_config()
assert config["output_dim"] == 2
assert config["bias"] is False
assert config["act"] == "relu"
def test_meanpool_agg_zero_neighbours():
agg = MeanPoolingAggregator(4, bias=False, act="linear", kernel_initializer="ones")
inp1 = keras.Input(shape=(1, 2))
inp2 = keras.Input(shape=(1, 0, 2))
out = agg([inp1, inp2])
# Now we have an input shape with a 0, the attention model switches to
# a MLP and the first group will have non-zero output size.
assert agg.weight_dims == [4, 0]
model = keras.Model(inputs=[inp1, inp2], outputs=out)
x1 = np.array([[[1, 1]]])
x2 = np.zeros((1, 1, 0, 2))
actual = model.predict([x1, x2])
expected = np.array([[[2, 2, 2, 2]]])
assert expected == pytest.approx(actual)
def test_meanpool_agg_constructor_1():
agg = MeanPoolingAggregator(output_dim=4, bias=True, act=lambda x: x + 1)
assert agg.output_dim == 4
assert agg.hidden_dim == 4
assert agg.has_bias
assert agg.act(2) == 3