def test_graphsage_apply_1():
gs = GraphSAGE(
layer_sizes=[2, 2, 2],
n_samples=[2, 2, 2],
bias=False,
input_dim=2,
multiplicity=1,
normalize=None,
kernel_initializer="ones",
)
inp = [keras.Input(shape=(i, 2)) for i in [1, 2, 4, 8]]
out = gs(inp)
model = keras.Model(inputs=inp, outputs=out)
x = [
np.array([[[1, 1]]]),
np.array([[[2, 2], [2, 2]]]),
np.array([[[3, 3], [3, 3], [3, 3], [3, 3]]]),
np.array([[[4, 4], [4, 4], [4, 4], [4, 4], [5, 5], [5, 5], [5, 5],
[5, 5]]]),
]
expected = np.array([[16, 25]])
actual = model.predict(x)
assert expected == pytest.approx(actual)
# Use the node model:
xinp, xout = gs.build()
model2 = keras.Model(inputs=xinp, outputs=xout)
assert pytest.approx(expected) == model2.predict(x)
def test_graphsage_passing_regularisers():
with pytest.raises(ValueError):
GraphSAGE(layer_sizes=[4],
n_samples=[2],
input_dim=2,
kernel_initializer="fred")
GraphSAGE(layer_sizes=[4],
n_samples=[2],
input_dim=2,
kernel_initializer="ones")
GraphSAGE(
layer_sizes=[4],
n_samples=[2],
input_dim=2,
kernel_initializer=initializers.ones(),
)
GraphSAGE(
layer_sizes=[4],
n_samples=[2],
input_dim=2,
kernel_regularizer=regularizers.l2(0.01),
)
with pytest.raises(ValueError):
GraphSAGE(layer_sizes=[4],
n_samples=[2],
input_dim=2,
kernel_regularizer="wilma")
def test_graphsage_constructor_passing_aggregator():
gs = GraphSAGE(layer_sizes=[4],
n_samples=[2],
input_dim=2,
aggregator=MeanAggregator)
assert gs.dims == [2, 4]
assert gs.n_samples == [2]
assert gs.max_hops == 1
assert gs.bias
assert len(gs._aggs) == 1
with pytest.raises(TypeError):
GraphSAGE(layer_sizes=[4], n_samples=[2], input_dim=2, aggregator=1)
def test_graphsage_serialize():
gs = GraphSAGE(
layer_sizes=[4],
n_samples=[2],
bias=False,
input_dim=2,
multiplicity=1,
normalize=None,
)
inp1 = keras.Input(shape=(1, 2))
inp2 = keras.Input(shape=(2, 2))
out = gs([inp1, inp2])
model = keras.Model(inputs=[inp1, inp2], outputs=out)
# Save model
model_json = model.to_json()
# Set all weights to one
model_weights = [np.ones_like(w) for w in model.get_weights()]
# Load model from json & set all weights
model2 = keras.models.model_from_json(
model_json, custom_objects={"MeanAggregator": MeanAggregator})
model2.set_weights(model_weights)
# Test loaded model
x1 = np.array([[[1, 1]]])
x2 = np.array([[[2, 2], [3, 3]]])
expected = np.array([[2, 2, 5, 5]])
actual = model2.predict([x1, x2])
assert expected == pytest.approx(actual)
def gs_nai_model(num_samples, generator, targets, optimizer, bias, dropout, normalize):
layer_sizes = [50] * len(num_samples)
graphsage = GraphSAGE(
layer_sizes=layer_sizes,
generator=generator,
bias=bias,
dropout=dropout,
normalize=normalize,
)
# Build the model and expose input and output sockets of graphsage, for node pair inputs:
x_inp, x_out = graphsage.build()
pred = tf.keras.layers.Dense(units=targets.shape[1], activation="softmax")(x_out)
model = tf.keras.Model(inputs=x_inp, outputs=pred)
model.compile(optimizer=optimizer, loss=tf.keras.losses.categorical_crossentropy)
return model
def test_kernel_and_bias_defaults():
gs = GraphSAGE(layer_sizes=[4, 4], n_samples=[2, 2], input_dim=2, multiplicity=1)
for layer in gs._aggs:
assert isinstance(layer.kernel_initializer, tf.initializers.GlorotUniform)
assert isinstance(layer.bias_initializer, tf.initializers.Zeros)
assert layer.kernel_regularizer is None
assert layer.bias_regularizer is None
assert layer.kernel_constraint is None
assert layer.bias_constraint is None
def unsup_gs_model(num_samples, generator, optimizer, bias, dropout, normalize):
layer_sizes = [50] * len(num_samples)
graphsage = GraphSAGE(
layer_sizes=layer_sizes,
generator=generator,
bias=bias,
dropout=dropout,
normalize=normalize,
)
# Build the model and expose input and output sockets of graphsage, for node pair inputs:
x_inp, x_out = graphsage.build()
prediction = link_classification(
output_dim=1, output_act="sigmoid", edge_embedding_method="ip"
)(x_out)
model = tf.keras.Model(inputs=x_inp, outputs=prediction)
model.compile(optimizer=optimizer, loss=tf.keras.losses.binary_crossentropy)
return model
def test_graphsage_constructor_1():
gs = GraphSAGE(layer_sizes=[4, 6, 8],
n_samples=[2, 4, 6],
input_dim=2,
bias=True,
dropout=0.5)
assert gs.dims == [2, 4, 6, 8]
assert gs.n_samples == [2, 4, 6]
assert gs.max_hops == 3
assert gs.bias
assert len(gs._aggs) == 3
def test_graphsage_apply():
gs = GraphSAGE(
layer_sizes=[4],
n_samples=[2],
bias=False,
input_dim=2,
normalize=None,
kernel_initializer="ones",
)
inp1 = keras.Input(shape=(1, 2))
inp2 = keras.Input(shape=(2, 2))
out = gs([inp1, inp2])
model = keras.Model(inputs=[inp1, inp2], outputs=out)
def test_graphsage_passing_activations():
gs = GraphSAGE(layer_sizes=[4], n_samples=[2], input_dim=2, multiplicity=1)
assert gs.activations == ["linear"]
gs = GraphSAGE(layer_sizes=[4, 4],
n_samples=[2, 2],
input_dim=2,
multiplicity=1)
assert gs.activations == ["relu", "linear"]
gs = GraphSAGE(layer_sizes=[4, 4, 4],
n_samples=[2, 2, 2],
input_dim=2,
multiplicity=1)
assert gs.activations == ["relu", "relu", "linear"]
with pytest.raises(ValueError):
GraphSAGE(
layer_sizes=[4, 4, 4],
n_samples=[2, 2, 2],
input_dim=2,
multiplicity=1,
activations=["relu"],
)
with pytest.raises(ValueError):
GraphSAGE(
layer_sizes=[4, 4, 4],
n_samples=[2, 2, 2],
input_dim=2,
multiplicity=1,
activations=["relu"] * 2,
)
with pytest.raises(ValueError):
GraphSAGE(
layer_sizes=[4, 4, 4],
n_samples=[2, 2, 2],
input_dim=2,
multiplicity=1,
activations=["fred", "wilma", "barney"],
)
gs = GraphSAGE(
layer_sizes=[4, 4, 4],
n_samples=[2, 2, 2],
input_dim=2,
multiplicity=1,
activations=["linear"] * 3,
)
assert gs.activations == ["linear"] * 3
def test_graphsage_constructor():
gs = GraphSAGE(layer_sizes=[4], n_samples=[2], input_dim=2, normalize="l2")
assert gs.dims == [2, 4]
assert gs.n_samples == [2]
assert gs.max_hops == 1
assert gs.bias
assert len(gs._aggs) == 1
# Check incorrect normalization flag
with pytest.raises(ValueError):
GraphSAGE(layer_sizes=[4],
n_samples=[2],
input_dim=2,
normalize=lambda x: x)
with pytest.raises(ValueError):
GraphSAGE(layer_sizes=[4],
n_samples=[2],
input_dim=2,
normalize="unknown")
# Check requirement for generator or n_samples
with pytest.raises(ValueError):
GraphSAGE(layer_sizes=[4])
# Construction from generator
G = example_graph_1(feature_size=3)
gen = GraphSAGENodeGenerator(G, batch_size=2, num_samples=[2,
2]).flow([1, 2])
gs = GraphSAGE(layer_sizes=[4, 8], generator=gen, bias=True)
assert gs.dims == [3, 4, 8]
assert gs.n_samples == [2, 2]
assert gs.max_hops == 2
assert gs.bias
assert len(gs._aggs) == 2
def test_graphsage_zero_neighbours():
gs = GraphSAGE(
layer_sizes=[2, 2],
n_samples=[0, 0],
bias=False,
input_dim=2,
normalize="none",
kernel_initializer="ones",
)
inp = [keras.Input(shape=(i, 2)) for i in [1, 0, 0]]
out = gs(inp)
model = keras.Model(inputs=inp, outputs=out)
x = [np.array([[[1.5, 1]]]), np.zeros((1, 0, 2)), np.zeros((1, 0, 2))]
actual = model.predict(x)
expected = np.array([[5, 5]])
assert actual == pytest.approx(expected)
def test_graphsage_constructor():
gs = GraphSAGE(layer_sizes=[4],
n_samples=[2],
input_dim=2,
normalize="l2",
multiplicity=1)
assert gs.dims == [2, 4]
assert gs.n_samples == [2]
assert gs.max_hops == 1
assert gs.bias
assert len(gs._aggs) == 1
# Check incorrect normalization flag
with pytest.raises(ValueError):
GraphSAGE(
layer_sizes=[4],
n_samples=[2],
input_dim=2,
normalize=lambda x: x,
multiplicity=1,
)
with pytest.raises(ValueError):
GraphSAGE(
layer_sizes=[4],
n_samples=[2],
input_dim=2,
normalize="unknown",
multiplicity=1,
)
# Check requirement for generator or n_samples
with pytest.raises(KeyError):
GraphSAGE(layer_sizes=[4])
# Construction from generator
G = example_graph(feature_size=3)
gen = GraphSAGENodeGenerator(G, batch_size=2, num_samples=[2, 2])
gs = GraphSAGE(layer_sizes=[4, 8], generator=gen, bias=True)
# The GraphSAGE should no longer accept a Sequence
t_gen = gen.flow([1, 2])
with pytest.raises(TypeError):
gs = GraphSAGE(layer_sizes=[4, 8], generator=t_gen, bias=True)
assert gs.dims == [3, 4, 8]
assert gs.n_samples == [2, 2]
assert gs.max_hops == 2
assert gs.bias
assert len(gs._aggs) == 2
def test_graphsage_save_load(tmpdir):
gs = GraphSAGE(layer_sizes=[4, 4], n_samples=[2, 2], input_dim=2, multiplicity=1)
test_utils.model_save_load(tmpdir, gs)