def test_ip(self):
""" Test the 'ip' binary operator on orthogonal vectors"""
x_src = np.random.randn(self.d)
x_src /= np.linalg.norm(x_src) # normalize x_src
x_dst = np.random.randn(self.d)
x_dst -= x_dst.dot(x_src) * x_src # make x_dst orthogonal to x_src
x_dst /= np.linalg.norm(x_dst) # normalize x_dst
expected = np.dot(x_src, x_dst)
sess = tf.InteractiveSession()
x_src = tf.constant(x_src, shape=(1, self.d), dtype="float64")
x_dst = tf.constant(x_dst, shape=(1, self.d), dtype="float64")
li = link_inference(edge_feature_method="ip")([x_src, x_dst])
print(
"link inference with 'ip' operator on orthonormal vectors: {}, expected: {}"
.format(li.eval(), expected))
assert li.eval() == pytest.approx(expected)
assert li.eval() == pytest.approx(0)
li = link_inference(edge_feature_method="ip")([x_src, x_src])
print("link inference with 'ip' operator on unit vector: ", li.eval())
assert li.eval() == pytest.approx(1)
sess.close()
def test_ip(self):
""" Test the 'ip' binary operator on orthogonal vectors"""
x_src, x_dst = make_orthonormal_vectors(self.d)
x_src = tf.constant(x_src, shape=(1, self.d), dtype="float64")
x_dst = tf.constant(x_dst, shape=(1, self.d), dtype="float64")
li = link_inference(edge_embedding_method="ip",
output_act="linear")([x_src, x_dst])
print("link inference with 'ip' operator on orthonormal vectors: {}".
format(li))
assert li.numpy() == pytest.approx(0, abs=1.5e-7)
li = link_inference(edge_embedding_method="ip",
output_act="linear")([x_src, x_src])
print("link inference with 'ip' operator on unit vector: ", li)
assert li.numpy() == pytest.approx(1, abs=1.5e-7)
# Test sigmoid activation
li = link_classification(edge_embedding_method="ip",
output_act="sigmoid")([x_src, x_dst])
assert li.numpy() == pytest.approx(0.5, abs=1.5e-7)
li = link_classification(edge_embedding_method="ip",
output_act="sigmoid")([x_src, x_src])
assert li.numpy() == pytest.approx(0.7310586, abs=1.5e-7)
def test_mul_l1_l2_avg(self):
""" Test the binary operators: 'mul'/'hadamard', 'l1', 'l2', 'avg'"""
x_src = np.random.randn(self.d)
x_src /= np.linalg.norm(x_src) # normalize x_src
x_dst = np.random.randn(self.d)
x_dst -= x_dst.dot(x_src) * x_src # make x_dst orthogonal to x_src
x_dst /= np.linalg.norm(x_dst) # normalize x_dst
x_src = x_src.reshape(1, 1, self.d)
x_dst = x_dst.reshape(1, 1, self.d)
inp_src = keras.Input(shape=(1, self.d))
inp_dst = keras.Input(shape=(1, self.d))
for op in ["mul", "l1", "l2", "avg"]:
out = link_inference(output_dim=self.d_out,
edge_embedding_method=op)([inp_src, inp_dst])
li = keras.Model(inputs=[inp_src, inp_dst], outputs=out)
res = li.predict(x=[x_src, x_dst])
print("link inference with '{}' operator: {}".format(
op, res.flatten()))
assert res.shape == (1, self.d_out)
assert isinstance(res.flatten()[0], np.float32)
def test_mul_l1_l2_avg(self):
""" Test the binary operators: 'mul'/'hadamard', 'l1', 'l2', 'avg'"""
x_src, x_dst = make_orthonormal_vectors(self.d)
x_src = x_src.reshape(1, 1, self.d)
x_dst = x_dst.reshape(1, 1, self.d)
inp_src = keras.Input(shape=(1, self.d))
inp_dst = keras.Input(shape=(1, self.d))
for op in ["mul", "l1", "l2", "avg", "concat"]:
out = link_inference(output_dim=self.d_out,
edge_embedding_method=op)([inp_src, inp_dst])
li = keras.Model(inputs=[inp_src, inp_dst], outputs=out)
print(x_src.shape)
res = li.predict(x=[x_src, x_dst])
print("link inference with '{}' operator: {}".format(
op, res.flatten()))
assert res.shape == (1, self.d_out)
assert isinstance(res.flatten()[0], np.float32)