def test_Ensemble_fit_generator():
tf.keras.backend.clear_session()
graph = example_graph_1(feature_size=10)
# base_model, keras_model, generator, train_gen
gnn_models = [
create_graphSAGE_model(graph),
create_HinSAGE_model(graph),
create_GCN_model(graph),
create_GAT_model(graph),
]
for gnn_model in gnn_models:
keras_model = gnn_model[1]
generator = gnn_model[2]
train_gen = gnn_model[3]
ens = Ensemble(keras_model, n_estimators=2, n_predictions=1)
ens.compile(optimizer=Adam(),
loss=categorical_crossentropy,
weighted_metrics=["acc"])
ens.fit_generator(train_gen, epochs=1, verbose=0, shuffle=False)
with pytest.raises(ValueError):
ens.fit_generator(
generator=generator, # wrong type
epochs=10,
validation_data=train_gen,
verbose=0,
shuffle=False,
)
def test_evaluate_generator_link_prediction():
edge_ids_test = np.array([[1, 2], [2, 3], [1, 3]])
edge_labels_test = np.array([1, 1, 0])
graph = example_graph_1(feature_size=4)
# base_model, keras_model, generator, train_gen
gnn_models = [
create_graphSAGE_model(graph, link_prediction=True),
create_HinSAGE_model(graph, link_prediction=True),
]
for gnn_model in gnn_models:
keras_model = gnn_model[1]
generator = gnn_model[2]
ens = Ensemble(keras_model, n_estimators=2, n_predictions=3)
ens.compile(optimizer=Adam(),
loss=binary_crossentropy,
weighted_metrics=["acc"])
# Check that passing invalid parameters is handled correctly. We will not check error handling for those
# parameters that Keras will be responsible for.
with pytest.raises(ValueError):
ens.evaluate_generator(
generator=generator,
test_data=edge_ids_test,
test_targets=edge_labels_test,
)
with pytest.raises(ValueError):
ens.evaluate_generator(
generator=generator,
test_data=edge_labels_test,
test_targets=None, # must give test_targets
)
with pytest.raises(ValueError):
ens.evaluate_generator(
generator=generator.flow(edge_ids_test, edge_labels_test),
test_data=edge_ids_test,
test_targets=edge_labels_test,
)
# We won't train the model instead use the initial random weights to test
# the evaluate_generator method.
test_metrics_mean, test_metrics_std = ens.evaluate_generator(
generator.flow(edge_ids_test, edge_labels_test))
assert len(test_metrics_mean) == len(test_metrics_std)
assert len(test_metrics_mean.shape) == 1
assert len(test_metrics_std.shape) == 1
def test_predict_generator():
# test_data = np.array([[0, 0], [1, 1], [0.8, 0.8]])
test_data = np.array([4, 5, 6])
test_targets = np.array([[1, 0], [0, 1], [0, 1]])
graph = example_graph_1(feature_size=2)
# base_model, keras_model, generator, train_gen
gnn_models = [
create_graphSAGE_model(graph),
create_HinSAGE_model(graph),
create_GCN_model(graph),
create_GAT_model(graph),
]
for i, gnn_model in enumerate(gnn_models):
keras_model = gnn_model[1]
generator = gnn_model[2]
ens = Ensemble(keras_model, n_estimators=2, n_predictions=2)
ens.compile(optimizer=Adam(),
loss=categorical_crossentropy,
weighted_metrics=["acc"])
test_gen = generator.flow(test_data)
# Check that passing invalid parameters is handled correctly. We will not check error handling for those
# parameters that Keras will be responsible for.
with pytest.raises(ValueError):
ens.predict_generator(generator=test_gen, predict_data=test_data)
# We won't train the model instead use the initial random weights to test
# the evaluate_generator method.
test_predictions = ens.predict_generator(test_gen, summarise=True)
print("test_predictions shape {}".format(test_predictions.shape))
if i > 1:
# GAT and GCN are full batch so we get a prediction for each node in the graph
assert len(test_predictions) == 6
else:
assert len(test_predictions) == len(test_data)
assert test_predictions.shape[1] == test_targets.shape[1]
test_predictions = ens.predict_generator(test_gen, summarise=False)
assert test_predictions.shape[0] == ens.n_estimators
assert test_predictions.shape[1] == ens.n_predictions
if i > 1:
assert test_predictions.shape[2] == 6
else:
assert test_predictions.shape[2] == len(test_data)
assert test_predictions.shape[3] == test_targets.shape[1]
def test_evaluate_generator():
test_data = np.array([3, 4, 5])
test_targets = np.array([[1, 0], [0, 1], [0, 1]])
graph = example_graph_1(feature_size=5)
# base_model, keras_model, generator, train_gen
gnn_models = [
create_graphSAGE_model(graph),
create_HinSAGE_model(graph),
create_GCN_model(graph),
create_GAT_model(graph),
]
for gnn_model in gnn_models:
keras_model = gnn_model[1]
generator = gnn_model[2]
ens = Ensemble(keras_model, n_estimators=2, n_predictions=1)
ens.compile(optimizer=Adam(),
loss=categorical_crossentropy,
weighted_metrics=["acc"])
# Check that passing invalid parameters is handled correctly. We will not check error handling for those
# parameters that Keras will be responsible for.
with pytest.raises(ValueError):
ens.evaluate_generator(generator=generator,
test_data=test_data,
test_targets=test_targets)
with pytest.raises(ValueError):
ens.evaluate_generator(
generator=generator,
test_data=test_data,
test_targets=None, # must give test_targets
)
with pytest.raises(ValueError):
ens.evaluate_generator(
generator=generator.flow(test_data, test_targets),
test_data=test_data,
test_targets=test_targets,
)
# We won't train the model instead use the initial random weights to test
# the evaluate_generator method.
test_metrics_mean, test_metrics_std = ens.evaluate_generator(
generator.flow(test_data, test_targets))
assert len(test_metrics_mean) == len(test_metrics_std)
assert len(test_metrics_mean.shape) == 1
assert len(test_metrics_std.shape) == 1
def test_predict_generator_link_prediction():
edge_ids_test = np.array([[1, 2], [2, 3], [1, 3]])
graph = example_graph_1(feature_size=2)
# base_model, keras_model, generator, train_gen
gnn_models = [
create_graphSAGE_model(graph, link_prediction=True),
create_HinSAGE_model(graph, link_prediction=True),
]
for gnn_model in gnn_models:
keras_model = gnn_model[1]
generator = gnn_model[2]
ens = Ensemble(keras_model, n_estimators=2, n_predictions=3)
ens.compile(optimizer=Adam(),
loss=binary_crossentropy,
weighted_metrics=["acc"])
test_gen = generator.flow(edge_ids_test)
# Check that passing invalid parameters is handled correctly. We will not check error handling for those
# parameters that Keras will be responsible for.
with pytest.raises(ValueError):
ens.predict_generator(generator=test_gen,
predict_data=edge_ids_test)
# We won't train the model instead use the initial random weights to test
# the evaluate_generator method.
test_predictions = ens.predict_generator(test_gen, summarise=True)
print("test_predictions shape {}".format(test_predictions.shape))
assert len(test_predictions) == len(edge_ids_test)
assert test_predictions.shape[1] == 1
test_predictions = ens.predict_generator(test_gen, summarise=False)
assert test_predictions.shape[0] == ens.n_estimators
assert test_predictions.shape[1] == ens.n_predictions
assert test_predictions.shape[2] == len(edge_ids_test)
assert test_predictions.shape[3] == 1
def test_compile():
graph = example_graph_1(feature_size=10)
# base_model, keras_model, generator, train_gen
gnn_models = [
create_graphSAGE_model(graph),
create_HinSAGE_model(graph),
create_graphSAGE_model(graph, link_prediction=True),
create_HinSAGE_model(graph, link_prediction=True),
create_GCN_model(graph),
create_GAT_model(graph),
]
for gnn_model in gnn_models:
keras_model = gnn_model[1]
ens = Ensemble(keras_model, n_estimators=2, n_predictions=5)
# These are actually raised by keras but I added a check just to make sure
with pytest.raises(ValueError):
ens.compile(optimizer=Adam(), loss=None, weighted_metrics=["acc"])
with pytest.raises(ValueError): # must specify the optimizer to use
ens.compile(optimizer=None,
loss=categorical_crossentropy,
weighted_metrics=["acc"])
with pytest.raises(
ValueError
): # The metric is made up so it should raise ValueError
ens.compile(
optimizer=Adam(),
loss=categorical_crossentropy,
weighted_metrics=["f1_accuracy"],
)
def test_fit_generator():
train_data = np.array([1, 2])
train_targets = np.array([[1, 0], [0, 1]])
graph = example_graph_1(feature_size=10)
# base_model, keras_model, generator, train_gen
gnn_models = [
create_graphSAGE_model(graph),
create_HinSAGE_model(graph),
create_graphSAGE_model(graph, link_prediction=True),
create_HinSAGE_model(graph, link_prediction=True),
create_GCN_model(graph),
create_GAT_model(graph),
]
for gnn_model in gnn_models:
keras_model = gnn_model[1]
generator = gnn_model[2]
train_gen = gnn_model[3]
ens = Ensemble(keras_model, n_estimators=2, n_predictions=1)
ens.compile(optimizer=Adam(),
loss=categorical_crossentropy,
weighted_metrics=["acc"])
# Specifying train_data and train_targets, implies the use of bagging so train_gen would
# be of the wrong type for this call to fit_generator.
with pytest.raises(ValueError):
ens.fit_generator(
train_gen,
train_data=train_data,
train_targets=train_targets,
epochs=10,
validation_generator=train_gen,
verbose=0,
shuffle=False,
)
with pytest.raises(ValueError):
ens.fit_generator(
generator=generator,
train_data=train_data,
train_targets=None, # Should not be None
epochs=10,
validation_generator=train_gen,
verbose=0,
shuffle=False,
)
with pytest.raises(ValueError):
ens.fit_generator(
generator=generator,
train_data=None,
train_targets=None,
epochs=10,
validation_generator=None,
verbose=0,
shuffle=False,
)
with pytest.raises(ValueError):
ens.fit_generator(
generator=generator,
train_data=train_data,
train_targets=train_targets,
epochs=10,
validation_generator=None,
verbose=0,
shuffle=False,
bag_size=
-1, # should be positive integer smaller than or equal to len(train_data) or None
)
with pytest.raises(ValueError):
ens.fit_generator(
generator=generator,
train_data=train_data,
train_targets=train_targets,
epochs=10,
validation_generator=None,
verbose=0,
shuffle=False,
bag_size=10, # larger than the number of training points
)