def test_attri2vec_serialize():
attri2vec = Attri2Vec(
layer_sizes=[4],
bias=False,
input_dim=2,
node_num=4,
multiplicity=2,
activation="linear",
normalize=None,
)
inp = keras.Input(shape=(2, ))
out = attri2vec(inp)
model = keras.Model(inputs=inp, 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)
model2.set_weights(model_weights)
# Test loaded model
x = np.array([[1, 2]])
expected = np.array([[3, 3, 3, 3]])
actual = model2.predict(x)
assert expected == pytest.approx(actual)
def test_attri2vec_apply():
attri2vec = Attri2Vec(
layer_sizes=[2, 2, 2],
bias=False,
input_dim=2,
node_num=4,
multiplicity=2,
activation="linear",
normalize=None,
)
x = np.array([[1, 2]])
expected = np.array([[12, 12]])
inp = keras.Input(shape=(2, ))
out = attri2vec(inp)
model1 = keras.Model(inputs=inp, outputs=out)
model_weights1 = [np.ones_like(w) for w in model1.get_weights()]
model1.set_weights(model_weights1)
actual = model1.predict(x)
assert expected == pytest.approx(actual)
# Use the node model:
xinp, xout = attri2vec.node_model()
model2 = keras.Model(inputs=xinp, outputs=xout)
model_weights2 = [np.ones_like(w) for w in model2.get_weights()]
model2.set_weights(model_weights2)
assert pytest.approx(expected) == model2.predict(x)
x1 = np.array([[3, 1]])
x2 = np.array([[2]])
y1 = np.array([[16, 16]])
y2 = np.array([[1, 1]])
# Test the build function:
xinp, xout = attri2vec.build()
model3 = keras.Model(inputs=xinp, outputs=xout)
model_weights3 = [np.ones_like(w) for w in model3.get_weights()]
model3.set_weights(model_weights3)
actual = model3.predict([x1, x2])
assert pytest.approx(y1) == actual[0]
assert pytest.approx(y2) == actual[1]
# Use the link model:
xinp, xout = attri2vec.link_model()
model4 = keras.Model(inputs=xinp, outputs=xout)
model_weights4 = [np.ones_like(w) for w in model4.get_weights()]
model4.set_weights(model_weights4)
actual = model4.predict([x1, x2])
assert pytest.approx(y1) == actual[0]
assert pytest.approx(y2) == actual[1]
def test_attri2vec_apply():
attri2vec = Attri2Vec(
layer_sizes=[2, 2, 2],
bias=False,
input_dim=2,
node_num=4,
multiplicity=2,
activation="linear",
normalize=None,
)
# the rest of the test assumes the weights are 1, to get the predictions to be easily computed,
# so let's build a basic model to set those weights (which are stored statefully in Attri2vec)
model = keras.Model(*attri2vec.in_out_tensors())
model.set_weights([np.ones_like(w) for w in model.get_weights()])
x = np.array([[1, 2]])
expected = np.array([[12, 12]])
inp = keras.Input(shape=(2, ))
out = attri2vec(inp)
model1 = keras.Model(inputs=inp, outputs=out)
actual = model1.predict(x)
assert expected == pytest.approx(actual)
# Use the node model:
xinp, xout = attri2vec.in_out_tensors(multiplicity=1)
model2 = keras.Model(inputs=xinp, outputs=xout)
assert pytest.approx(expected) == model2.predict(x)
x1 = np.array([[3, 1]])
x2 = np.array([[2]])
y1 = np.array([[16, 16]])
y2 = np.array([[1, 1]])
# Test the build function:
xinp, xout = attri2vec.in_out_tensors()
model3 = keras.Model(inputs=xinp, outputs=xout)
actual = model3.predict([x1, x2])
assert pytest.approx(y1) == actual[0]
assert pytest.approx(y2) == actual[1]
# Use the link model:
xinp, xout = attri2vec.in_out_tensors()
model4 = keras.Model(inputs=xinp, outputs=xout)
actual = model4.predict([x1, x2])
assert pytest.approx(y1) == actual[0]
assert pytest.approx(y2) == actual[1]