def test_rotate(knowledge_graph):
margin = 2.34
norm_order = 1.234
# this test creates a random untrained model and predicts every possible edge in the graph, and
# compares that to a direct implementation of the scoring method in the paper
gen = KGTripleGenerator(knowledge_graph, 3)
# use a random initializer with a large range, so that any differences are obvious
init = initializers.RandomUniform(-1, 1)
rotate_model = RotatE(
gen, 5, margin=margin, norm_order=norm_order, embeddings_initializer=init
)
x_inp, x_out = rotate_model.in_out_tensors()
model = Model(x_inp, x_out)
model.compile(loss=tf_losses.BinaryCrossentropy(from_logits=True))
every_edge = itertools.product(
knowledge_graph.nodes(),
knowledge_graph._edges.types.pandas_index,
knowledge_graph.nodes(),
)
df = triple_df(*every_edge)
# check the model can be trained on a few (uneven) batches
model.fit(
gen.flow(df.iloc[:7], negative_samples=2),
validation_data=gen.flow(df.iloc[7:14], negative_samples=3),
)
# compute the exact values based on the model by extracting the embeddings for each element and
# doing the y_(e_1)^T M_r y_(e_2) = <e_1, w_r, e_2> inner product
s_idx = knowledge_graph.node_ids_to_ilocs(df.source)
r_idx = knowledge_graph._edges.types.to_iloc(df.label)
o_idx = knowledge_graph.node_ids_to_ilocs(df.target)
nodes, edge_types = rotate_model.embeddings()
# the rows correspond to the embeddings for the given edge, so we can do bulk operations
e_s = nodes[s_idx, :]
w_r = edge_types[r_idx, :]
e_o = nodes[o_idx, :]
# every edge-type embedding should be a unit rotation
np.testing.assert_allclose(np.abs(w_r), 1)
actual = margin - np.linalg.norm(e_s * w_r - e_o, ord=norm_order, axis=1)
# predict every edge using the model
prediction = model.predict(gen.flow(df))
# (use an absolute tolerance to allow for catastrophic cancellation around very small values)
np.testing.assert_allclose(prediction[:, 0], actual, rtol=1e-3, atol=1e-14)
# the model is stateful (i.e. it holds the weights permanently) so the predictions with a second
# 'build' should be the same as the original one
model2 = Model(*rotate_model.in_out_tensors())
prediction2 = model2.predict(gen.flow(df))
np.testing.assert_array_equal(prediction, prediction2)
def test_complex(knowledge_graph, sample_strategy):
# this test creates a random untrained model and predicts every possible edge in the graph, and
# compares that to a direct implementation of the scoring method in the paper
gen = KGTripleGenerator(knowledge_graph, 3)
# use a random initializer with a large positive range, so that any differences are obvious
init = initializers.RandomUniform(-1, 1)
complex_model = ComplEx(gen, 5, embeddings_initializer=init)
x_inp, x_out = complex_model.in_out_tensors()
model = Model(x_inp, x_out)
if sample_strategy == "uniform":
loss = tf_losses.BinaryCrossentropy(from_logits=True)
else:
loss = sg_losses.SelfAdversarialNegativeSampling()
model.compile(loss=loss)
every_edge = itertools.product(
knowledge_graph.nodes(),
knowledge_graph._edges.types.pandas_index,
knowledge_graph.nodes(),
)
df = triple_df(*every_edge)
# check the model can be trained on a few (uneven) batches
model.fit(
gen.flow(df.iloc[:7], negative_samples=2, sample_strategy=sample_strategy),
validation_data=gen.flow(
df.iloc[7:14], negative_samples=3, sample_strategy=sample_strategy
),
)
# compute the exact values based on the model by extracting the embeddings for each element and
# doing the Re(<e_s, w_r, conj(e_o)>) inner product
s_idx = knowledge_graph.node_ids_to_ilocs(df.source)
r_idx = knowledge_graph._edges.types.to_iloc(df.label)
o_idx = knowledge_graph.node_ids_to_ilocs(df.target)
nodes, edge_types = complex_model.embeddings()
# the rows correspond to the embeddings for the given edge, so we can do bulk operations
e_s = nodes[s_idx, :]
w_r = edge_types[r_idx, :]
e_o = nodes[o_idx, :]
actual = (e_s * w_r * e_o.conj()).sum(axis=1).real
# predict every edge using the model
prediction = model.predict(gen.flow(df))
# (use an absolute tolerance to allow for catastrophic cancellation around very small values)
np.testing.assert_allclose(prediction[:, 0], actual, rtol=1e-3, atol=1e-6)
# the model is stateful (i.e. it holds the weights permanently) so the predictions with a second
# 'build' should be the same as the original one
model2 = Model(*complex_model.in_out_tensors())
prediction2 = model2.predict(gen.flow(df))
np.testing.assert_array_equal(prediction, prediction2)
def test_kg_triple_generator(knowledge_graph):
gen = KGTripleGenerator(knowledge_graph, 2)
edges = triple_df(("a", "W", "b"), ("c", "X", "a"), ("d", "Y", "c"))
seq = gen.flow(edges)
check_sequence_output(seq[0], 2, None)
check_sequence_output(seq[1], 1, None)
seq = gen.flow(edges, negative_samples=10)
check_sequence_output(seq[0], 2, 10, knowledge_graph.number_of_nodes())
def test_complex(knowledge_graph):
# this test creates a random untrained model and predicts every possible edge in the graph, and
# compares that to a direct implementation of the scoring method in the paper
gen = KGTripleGenerator(knowledge_graph, 3)
# use a random initializer with a large positive range, so that any differences are obvious
init = initializers.RandomUniform(-1, 1)
x_inp, x_out = ComplEx(gen, 5, embeddings_initializer=init).build()
model = Model(x_inp, x_out)
model.compile(loss=losses.BinaryCrossentropy(from_logits=True))
every_edge = itertools.product(
knowledge_graph.nodes(),
knowledge_graph._edges.types.pandas_index,
knowledge_graph.nodes(),
)
df = triple_df(*every_edge)
# check the model can be trained on a few (uneven) batches
model.fit(
gen.flow(df.iloc[:7], negative_samples=2),
validation_data=gen.flow(df.iloc[7:14], negative_samples=3),
)
# compute the exact values based on the model by extracting the embeddings for each element and
# doing the Re(<e_s, w_r, conj(e_o)>) inner product
s_idx = knowledge_graph._get_index_for_nodes(df.source)
r_idx = knowledge_graph._edges.types.to_iloc(df.label)
o_idx = knowledge_graph._get_index_for_nodes(df.target)
nodes, edge_types = ComplEx.embeddings(model)
# the rows correspond to the embeddings for the given edge, so we can do bulk operations
e_s = nodes[s_idx, :]
w_r = edge_types[r_idx, :]
e_o = nodes[o_idx, :]
actual = (e_s * w_r * e_o.conj()).sum(axis=1).real
# predict every edge using the model
prediction = model.predict(gen.flow(df))
# (use an absolute tolerance to allow for catastrophic cancellation around very small values)
assert np.allclose(prediction[:, 0], actual, rtol=1e-3, atol=1e-14)
def test_model_rankings(model_maker):
nodes = pd.DataFrame(index=["a", "b", "c", "d"])
rels = ["W", "X", "Y", "Z"]
empty = pd.DataFrame(columns=["source", "target"])
every_edge = itertools.product(nodes.index, rels, nodes.index)
every_edge_df = triple_df(*every_edge)
no_edges = StellarDiGraph(nodes, {name: empty for name in rels})
# the filtering is most interesting when there's a smattering of edges, somewhere between none
# and all; this does a stratified sample by label, to make sure there's at least one edge from
# each label.
one_per_label_df = (every_edge_df.groupby("label").apply(
lambda df: df.sample(n=1)).droplevel(0))
others_df = every_edge_df.sample(frac=0.25)
some_edges_df = pd.concat([one_per_label_df, others_df], ignore_index=True)
some_edges = StellarDiGraph(
nodes,
{
name: df.drop(columns="label")
for name, df in some_edges_df.groupby("label")
},
)
all_edges = StellarDiGraph(
nodes=nodes,
edges={
name: df.drop(columns="label")
for name, df in every_edge_df.groupby("label")
},
)
gen = KGTripleGenerator(all_edges, 3)
sg_model = model_maker(gen, embedding_dimension=5)
x_inp, x_out = sg_model.in_out_tensors()
model = Model(x_inp, x_out)
raw_some, filtered_some = sg_model.rank_edges_against_all_nodes(
gen.flow(every_edge_df), some_edges)
# basic check that the ranks are formed correctly
assert raw_some.dtype == int
assert np.all(raw_some >= 1)
# filtered ranks are never greater, and sometimes less
assert np.all(filtered_some <= raw_some)
assert np.any(filtered_some < raw_some)
raw_no, filtered_no = sg_model.rank_edges_against_all_nodes(
gen.flow(every_edge_df), no_edges)
np.testing.assert_array_equal(raw_no, raw_some)
# with no edges, filtering does nothing
np.testing.assert_array_equal(raw_no, filtered_no)
raw_all, filtered_all = sg_model.rank_edges_against_all_nodes(
gen.flow(every_edge_df), all_edges)
np.testing.assert_array_equal(raw_all, raw_some)
# when every edge is known, the filtering should eliminate every possibility
assert np.all(filtered_all == 1)
# check the ranks against computing them from the model predictions directly. That is, for each
# edge, compare the rank against one computed by counting the predictions. This computes the
# filtered ranks naively too.
predictions = model.predict(gen.flow(every_edge_df))
for (source, rel, target), score, raw, filtered in zip(
every_edge_df.itertuples(index=False), predictions, raw_some,
filtered_some):
# rank for the subset specified by the given selector
def rank(compare_selector):
return 1 + (predictions[compare_selector] > score).sum()
same_r = every_edge_df.label == rel
same_s_r = (every_edge_df.source == source) & same_r
expected_raw_mod_o_rank = rank(same_s_r)
assert raw[0] == expected_raw_mod_o_rank
known_objects = some_edges_df[(some_edges_df.source == source)
& (some_edges_df.label == rel)]
object_is_unknown = ~every_edge_df.target.isin(known_objects.target)
expected_filt_mod_o_rank = rank(same_s_r & object_is_unknown)
assert filtered[0] == expected_filt_mod_o_rank
same_r_o = same_r & (every_edge_df.target == target)
expected_raw_mod_s_rank = rank(same_r_o)
assert raw[1] == expected_raw_mod_s_rank
known_subjects = some_edges_df[(some_edges_df.label == rel)
& (some_edges_df.target == target)]
subject_is_unknown = ~every_edge_df.source.isin(known_subjects.source)
expected_filt_mod_s_rank = rank(subject_is_unknown & same_r_o)
assert filtered[1] == expected_filt_mod_s_rank
def test_rote_roth(knowledge_graph, model_class):
# this test creates a random untrained model and predicts every possible edge in the graph, and
# compares that to a direct implementation of the scoring method in the paper
gen = KGTripleGenerator(knowledge_graph, 3)
# use a random initializer with a large range, so that any differences are obvious
init = initializers.RandomUniform(-1, 1)
rot_model = model_class(gen, 6, embeddings_initializer=init)
x_inp, x_out = rot_model.in_out_tensors()
model = Model(x_inp, x_out)
model.summary()
model.compile(loss=tf_losses.BinaryCrossentropy(from_logits=True))
every_edge = itertools.product(
knowledge_graph.nodes(),
knowledge_graph._edges.types.pandas_index,
knowledge_graph.nodes(),
)
df = triple_df(*every_edge)
# check the model can be trained on a few (uneven) batches
model.fit(
gen.flow(df.iloc[:7], negative_samples=2),
validation_data=gen.flow(df.iloc[7:14], negative_samples=3),
)
# predict every edge using the model
prediction = model.predict(gen.flow(df))
(node_emb, node_bias), (et_emb, et_theta) = rot_model.embedding_arrays()
if model_class is RotE:
# compute the exact values based on the model, for RotationE (the RotationH model is too
# hard to test directly)
s_idx = knowledge_graph.node_ids_to_ilocs(df.source)
r_idx = knowledge_graph.edge_type_names_to_ilocs(df.label)
o_idx = knowledge_graph.node_ids_to_ilocs(df.target)
# the rows correspond to the embeddings for the given edge, so we can do bulk operations
e_s = node_emb[s_idx, :]
b_s = node_bias[s_idx, 0]
r_r = et_emb[r_idx, :]
theta_r = et_theta[r_idx, :]
e_o = node_emb[o_idx, :]
b_o = node_bias[o_idx, 0]
rot_r = np.cos(theta_r) + 1j * np.sin(theta_r)
assert e_s.dtype == np.float32
rotated = (e_s.view(np.complex64) * rot_r).view(np.float32)
actual = -np.linalg.norm(rotated + r_r - e_o, axis=-1) ** 2 + b_s + b_o
np.testing.assert_allclose(prediction[:, 0], actual, rtol=1e-3, atol=1e-14)
# the model is stateful (i.e. it holds the weights permanently) so the predictions with a second
# 'build' should be the same as the original one
model2 = Model(*rot_model.in_out_tensors())
prediction2 = model2.predict(gen.flow(df))
np.testing.assert_array_equal(prediction, prediction2)
def test_kg_triple_generator_errors(knowledge_graph):
gen = KGTripleGenerator(knowledge_graph, 10)
with pytest.raises(TypeError, match="edges: expected.*found int"):
gen.flow(1)
with pytest.raises(KeyError, match="fake"):
gen.flow(triple_df(("fake", "W", "b")))
with pytest.raises(KeyError, match="fake"):
gen.flow(triple_df(("a", "fake", "b")))
with pytest.raises(KeyError, match="fake"):
gen.flow(triple_df(("a", "W", "fake")))
with pytest.raises(TypeError,
match="negative_samples: expected.*found str"):
gen.flow(triple_df(), negative_samples="foo")
with pytest.raises(ValueError,
match="negative_samples: expected.*found -1"):
gen.flow(triple_df(), negative_samples=-1)
def test_kg_triple_generator_errors(knowledge_graph):
gen = KGTripleGenerator(knowledge_graph, 10)
with pytest.raises(TypeError, match="edges: expected.*found int"):
gen.flow(1)
with pytest.raises(KeyError, match="fake"):
gen.flow(triple_df(("fake", "W", "b")))
with pytest.raises(KeyError, match="fake"):
gen.flow(triple_df(("a", "fake", "b")))
with pytest.raises(KeyError, match="fake"):
gen.flow(triple_df(("a", "W", "fake")))
with pytest.raises(TypeError, match="negative_samples: expected.*found str"):
gen.flow(triple_df(), negative_samples="foo")
with pytest.raises(ValueError, match="negative_samples: expected.*found -1"):
gen.flow(triple_df(), negative_samples=-1)
with pytest.raises(
ValueError,
match="sample_strategy: expected one of 'uniform', 'self-adversarial', found None",
):
gen.flow(triple_df(), sample_strategy=None)
with pytest.raises(
ValueError, match="sample_strategy: expected .* found 'UNIFORM'"
):
# case-sensitive
gen.flow(triple_df(), sample_strategy="UNIFORM")