def test_WatchYourStep_bad_init(barbell):
generator = AdjacencyPowerGenerator(barbell, num_powers=5)
with pytest.raises(TypeError, match="num_walks: expected.* found float"):
wys = WatchYourStep(generator, num_walks=10.0)
with pytest.raises(ValueError, match="num_walks: expected.* found 0"):
wys = WatchYourStep(generator, num_walks=0)
with pytest.raises(TypeError, match="embedding_dimension: expected.* found float"):
wys = WatchYourStep(generator, embedding_dimension=10.0)
with pytest.raises(ValueError, match="embedding_dimension: expected.* found 1"):
wys = WatchYourStep(generator, embedding_dimension=1)
def test_WatchYourStep_init(barbell):
generator = AdjacencyPowerGenerator(barbell, num_powers=5)
wys = WatchYourStep(generator)
assert wys.num_powers == 5
assert wys.n_nodes == len(barbell.nodes())
assert wys.num_walks == 80
assert wys.embedding_dimension == 64
def test_WatchYourStep_embeddings(barbell):
generator = AdjacencyPowerGenerator(barbell, num_powers=5)
wys = WatchYourStep(generator, embeddings_initializer="ones")
x_in, x_out = wys.in_out_tensors()
model = Model(inputs=x_in, outputs=x_out)
model.compile(optimizer="adam", loss=graph_log_likelihood)
embs = wys.embeddings()
assert (embs == 1).all()
def test_WatchYourStep(barbell, weighted):
generator = AdjacencyPowerGenerator(barbell, num_powers=5, weighted=weighted)
gen = generator.flow(batch_size=4)
wys = WatchYourStep(generator)
x_in, x_out = wys.in_out_tensors()
model = Model(inputs=x_in, outputs=x_out)
model.compile(optimizer="adam", loss=graph_log_likelihood)
model.fit(gen, epochs=1, steps_per_epoch=int(len(barbell.nodes()) // 4))
embs = wys.embeddings()
assert embs.shape == (len(barbell.nodes()), wys.embedding_dimension)
# build() should always return tensors backed by the same trainable weights, and thus give the
# same predictions
preds1 = model.predict(gen, steps=8)
preds2 = Model(*wys.in_out_tensors()).predict(gen, steps=8)
np.testing.assert_array_equal(preds1, preds2)
def _dispatch_transductive_layer(generator, model_name, params):
"""Create an embedding layer."""
embedding_layer = None
if model_name == "watchyourstep":
embedding_layer = WatchYourStep(
generator,
num_walks=params["number_of_walks"],
embedding_dimension=params["embedding_dimension"],
attention_regularizer=regularizers.l2(0.5),
)
elif model_name == "distmult":
embedding_layer = DistMult(
generator,
embedding_dimension=params["embedding_dimension"],
embeddings_regularizer=regularizers.l2(1e-7),
)
elif model_name == "complex":
embedding_layer = ComplEx(
generator,
embedding_dimension=params["embedding_dimension"],
embeddings_regularizer=regularizers.l2(1e-7),
)
return embedding_layer
def test_WatchYourStep_save_load(tmpdir, barbell):
generator = AdjacencyPowerGenerator(barbell, num_powers=5)
wys = WatchYourStep(generator)
test_utils.model_save_load(tmpdir, wys)