def graphsage_pipeline(G, node_subjects, layer_sizes=[32, 32]):
train_subjects, val_subjects, test_subjects = training_split(node_subjects)
batch_size = 50
num_samples = [10, 5]
generator = GraphSAGENodeGenerator(G, batch_size, num_samples)
train_gen = generator.flow(train_subjects.index,
train_subjects.values,
shuffle=True)
graphsage_model = GraphSAGE(
layer_sizes=layer_sizes,
generator=generator,
bias=True,
dropout=0.5,
)
model = build_model(graphsage_model, train_subjects.values.shape[1])
val_gen = generator.flow(val_subjects.index, val_subjects.values)
es_callback = EarlyStopping(monitor="val_acc",
patience=50,
restore_best_weights=True)
history = model.fit(train_gen,
epochs=200,
validation_data=val_gen,
verbose=0,
shuffle=False,
callbacks=[es_callback])
plot_results(history)
test_metrics(generator, model, test_subjects)
def graph_node_classifier(name, train_data, layer_sizes=[32, 32], verbose=1):
"""
Build and return a neural node classification model.
Notes: Only mutually-exclusive class labels are supported.
Args:
name (string): one of:
- 'graphsage' for GraphSAGE model
(only GraphSAGE currently supported)
train_data (NodeSequenceWrapper): a deepwrap.graph.sg_wrappers.NodeSequenceWrapper object
verbose (boolean): verbosity of output
Return:
model (Model): A Keras Model instance
"""
from .sg_wrappers import NodeSequenceWrapper
# check argument
if not isinstance(train_data, NodeSequenceWrapper):
err = """
train_data must be a deepwrap.graph.sg_wrappers.NodeSequenceWrapper object
"""
raise Exception(err)
if len(layer_sizes) != 2:
raise ValueError('layer_sizes must be of length 2')
num_classes = U.nclasses_from_data(train_data)
# determine multilabel
multilabel = U.is_multilabel(train_data)
if multilabel:
raise ValueError(
'Multi-label classification not currently supported for graphs.')
U.vprint("Is Multi-Label? %s" % (multilabel), verbose=verbose)
# set loss and activations
loss_func = 'categorical_crossentropy'
activation = 'softmax'
# import stellargraph
try:
import stellargraph as sg
from stellargraph.layer import GraphSAGE
except:
raise Exception(SG_ERRMSG)
if version.parse(sg.__version__) < version.parse('0.8'):
raise Exception(SG_ERRMSG)
# build a GraphSAGE node classification model
graphsage_model = GraphSAGE(
layer_sizes=layer_sizes,
generator=train_data,
bias=True,
dropout=0.5,
)
# x_inp, x_out = graphsage_model.default_model(flatten_output=True)
x_inp, x_out = graphsage_model.build()
prediction = Dense(units=num_classes, activation=activation)(x_out)
model = Model(inputs=x_inp, outputs=prediction)
model.compile(optimizer='adam', loss=loss_func, metrics=["accuracy"])
U.vprint('done', verbose=verbose)
return model
def graph_link_predictor(name,
train_data,
preproc,
layer_sizes=[20, 20],
verbose=1):
"""
Build and return a neural link prediction model.
Args:
name (string): one of:
- 'graphsage' for GraphSAGE model
(only GraphSAGE currently supported)
train_data (LinkSequenceWrapper): a ktrain.graph.sg_wrappers.LinkSequenceWrapper object
preproc(LinkPreprocessor): a LinkPreprocessor instance
verbose (boolean): verbosity of output
Return:
model (Model): A Keras Model instance
"""
from .sg_wrappers import LinkSequenceWrapper
# check argument
if not isinstance(train_data, LinkSequenceWrapper):
err = """
train_data must be a ktrain.graph.sg_wrappers.LinkSequenceWrapper object
"""
raise Exception(err)
if len(layer_sizes) != len(preproc.sample_sizes):
raise ValueError(
'number of layer_sizes must match len(preproc.sample_sizes)')
num_classes = U.nclasses_from_data(train_data)
# set loss and activations
loss_func = 'categorical_crossentropy'
activation = 'softmax'
# import stellargraph
try:
import stellargraph as sg
from stellargraph.layer import GraphSAGE, link_classification
except:
raise Exception(SG_ERRMSG)
if version.parse(sg.__version__) < version.parse('0.8'):
raise Exception(SG_ERRMSG)
# build a GraphSAGE link prediction model
graphsage = GraphSAGE(layer_sizes=layer_sizes,
generator=train_data,
bias=True,
dropout=0.3)
x_inp, x_out = graphsage.build()
prediction = link_classification(output_dim=1,
output_act="relu",
edge_embedding_method='ip')(x_out)
model = Model(inputs=x_inp, outputs=prediction)
model.compile(optimizer=U.DEFAULT_OPT,
loss='binary_crossentropy',
metrics=["accuracy"])
return model
def initialize(self,**hyper_params):
if(not "batch_size" in hyper_params.keys()):
batch_size = 16
if(not "layer_sizes" in hyper_params.keys()):
num_samples = [25, 10]
if(not "num_samples" in hyper_params.keys()):
layer_sizes = [256, 256]
if(not "bias" in hyper_params.keys()):
bias = True
if(not "dropout" in hyper_params.keys()):
dropout = 0.0
if(not "lr" in hyper_params.keys()):
lr = 1e-3
if(not "num_walks" in hyper_params.keys()):
num_walks = 1
if(not "length" in hyper_params.keys()):
length = 5
self.graph = sg.StellarGraph(nodes=self.nodes_df,edges=self.edges_df)
self.nodes = list(self.graph.nodes())
del self.nodes_df
del self.edges_df
unsupervised_samples = UnsupervisedSampler(
self.graph, nodes=self.nodes, length=length, number_of_walks=num_walks
)
# Train iterators
train_gen = GraphSAGELinkGenerator(self.graph, batch_size, num_samples)
self.train_flow = train_gen.flow(unsupervised_samples)
# Model defining - Keras functional API + Stellargraph layers
graphsage = GraphSAGE(
layer_sizes=layer_sizes, generator=train_gen, bias=bias, dropout=dropout, normalize="l2"
)
x_inp, x_out = graphsage.in_out_tensors()
prediction = link_classification(
output_dim=1, output_act="sigmoid", edge_embedding_method="ip"
)(x_out)
self.model = keras.Model(inputs=x_inp, outputs=prediction)
self.model.compile(
optimizer=keras.optimizers.Adam(lr=lr),
loss=keras.losses.binary_crossentropy,
metrics=[keras.metrics.binary_accuracy],
)
x_inp_src = x_inp[0::2]
x_out_src = x_out[0]
self.embedding_model = keras.Model(inputs=x_inp_src, outputs=x_out_src)
self.node_gen = GraphSAGENodeGenerator(self.graph, batch_size, num_samples).flow(self.nodes)
return self.model.get_weights()
def create_graphsage(train_gen):
return GraphSAGE(
layer_sizes=config.LAYER_SIZES,
generator=train_gen,
bias=True,
dropout=config.DROPOUT,
normalize="l2",
)
def initialize(self,**hyper_params):
if(not "batch_size" in hyper_params.keys()):
batch_size = 20
if(not "layer_sizes" in hyper_params.keys()):
num_samples = [20, 10]
if(not "num_samples" in hyper_params.keys()):
layer_sizes = [10, 10 ]
if(not "bias" in hyper_params.keys()):
bias = True
if(not "dropout" in hyper_params.keys()):
dropout = 0.1
if(not "lr" in hyper_params.keys()):
lr = 1e-2
graph = sg.StellarGraph(nodes=self.nodes,edges=self.edges)
# Test split
edge_splitter_test = EdgeSplitter(graph)
self.graph_test, edge_ids_test, edge_labels_test = edge_splitter_test.train_test_split(
p=0.1, method="global", keep_connected=True, seed = 42
)
# Train split
edge_splitter_train = EdgeSplitter(self.graph_test)
self.graph_train, edge_ids_train, edge_labels_train = edge_splitter_train.train_test_split(
p=0.1, method="global", keep_connected=True, seed = 42
)
# Train iterators
train_gen = GraphSAGELinkGenerator(self.graph_train, batch_size, num_samples, seed = 42)
self.train_flow = train_gen.flow(edge_ids_train, edge_labels_train, shuffle=True)
# Test iterators
test_gen = GraphSAGELinkGenerator(self.graph_train, batch_size, num_samples, seed = 42)
self.test_flow = test_gen.flow(edge_ids_test, edge_labels_test, shuffle=True)
# Model defining - Keras functional API + Stellargraph layers
graphsage = GraphSAGE(
layer_sizes=layer_sizes, generator=train_gen, bias=bias, dropout=dropout
)
x_inp, x_out = graphsage.in_out_tensors()
prediction = link_classification(
output_dim=1, output_act="sigmoid", edge_embedding_method="ip"
)(x_out)
self.model = keras.Model(inputs=x_inp, outputs=prediction)
self.model.compile(
optimizer=keras.optimizers.Adam(lr=lr),
loss=keras.losses.binary_crossentropy,
metrics=[keras.metrics.BinaryAccuracy(),keras.metrics.Recall(),keras.metrics.AUC(),keras.metrics.Precision()],
)
# return number of training and testing examples
return edge_ids_train.shape[0],edge_ids_test.shape[0]
def train_clf(self, graph, L):
'''
Train GraphSage model with updated labeled pool L
Return new trained model
'''
train_targets = self.target_encoding.transform(
self.df_targets.loc[L].to_dict("records"))
train_gen = self.generator.flow(L, train_targets)
gsage = GraphSAGE(layer_sizes=[32, 32],
generator=self.generator,
bias=True,
dropout=0.5)
x_inp, x_out = gsage.build()
predictions = layers.Dense(units=train_targets.shape[1],
activation="softmax")(x_out)
class_support = dict(Counter(self.df_targets.loc[L]["label"]))
classes = sorted(self.data.class_labels)
counts = [
class_support[c] if c in class_support else 0 for c in classes
]
weights = np.sum(counts) / np.array(counts)
weighted_loss = self.weighted_categorical_crossentropy(weights)
model = Model(inputs=x_inp, outputs=predictions)
model.compile(
optimizer=optimizers.Adam(lr=0.2),
# loss=losses.categorical_crossentropy,
loss=weighted_loss,
metrics=["acc"],
)
# if not os.path.isdir("model_logs"):
# os.makedirs("model_logs")
# es_callback = EarlyStopping(
# monitor="acc", patience=50
# ) # patience is the number of epochs to wait before early stopping in case of no further improvement
# mc_callback = ModelCheckpoint(
# "model_logs/best_model.h5", monitor="acc", save_best_only=True, save_weights_only=True
# )
history = model.fit_generator(
train_gen,
epochs=50,
verbose=0,
shuffle=
False, # this should be False, since shuffling data means shuffling the whole graph
# callbacks=[es_callback, mc_callback],
)
# model.load_weights("model_logs/best_model.h5")
return model
def initialize(self, **hyper_params):
if (not "batch_size" in hyper_params.keys()):
batch_size = 20
if (not "layer_sizes" in hyper_params.keys()):
num_samples = [20, 10]
if (not "num_samples" in hyper_params.keys()):
layer_sizes = [20, 20]
if (not "bias" in hyper_params.keys()):
bias = True
if (not "dropout" in hyper_params.keys()):
dropout = 0.3
if (not "lr" in hyper_params.keys()):
lr = 1e-3
if (not "train_split" in hyper_params.keys()):
train_split = 0.2
self.graph = sg.StellarGraph(nodes=self.nodes, edges=self.edges)
# Train split
edge_splitter_train = EdgeSplitter(self.graph)
graph_train, edge_ids_train, edge_labels_train = edge_splitter_train.train_test_split(
p=train_split, method="global", keep_connected=True)
# Train iterators
train_gen = GraphSAGELinkGenerator(graph_train, batch_size,
num_samples)
self.train_flow = train_gen.flow(edge_ids_train,
edge_labels_train,
shuffle=True)
# Model defining - Keras functional API + Stellargraph layers
graphsage = GraphSAGE(layer_sizes=layer_sizes,
generator=train_gen,
bias=bias,
dropout=dropout)
x_inp, x_out = graphsage.in_out_tensors()
prediction = link_classification(output_dim=1,
output_act="relu",
edge_embedding_method="ip")(x_out)
self.model = keras.Model(inputs=x_inp, outputs=prediction)
self.model.compile(
optimizer=keras.optimizers.Adam(lr=lr),
loss=keras.losses.binary_crossentropy,
metrics=["acc"],
)
return self.model.get_weights()
def run_model(self):
graph_sampled, label_series_sampled = self.prepare_data_for_stellargraph(
)
train_targets, valid_targets, test_targets, train_labels, valid_labels, test_labels = self.get_train_valid_test(
label_series_sampled)
batch_size = self.hyperparams["batch_size"]
num_samples = self.hyperparams["num_samples"]
generator = GraphSAGENodeGenerator(graph_sampled, batch_size,
num_samples)
train_gen = generator.flow(train_labels.index,
train_targets,
shuffle=True)
graphsage_model = GraphSAGE(
layer_sizes=self.hyperparams["layer_sizes"],
generator=generator,
bias=self.hyperparams["bias"],
dropout=self.hyperparams["dropout"],
)
x_inp, x_out = graphsage_model.in_out_tensors()
prediction = layers.Dense(units=train_targets.shape[1],
activation="softmax")(x_out)
model = Model(inputs=x_inp, outputs=prediction)
model.compile(
optimizer=optimizers.Adam(lr=self.hyperparams["lr"]),
loss=losses.categorical_crossentropy,
metrics=["acc"],
)
valid_gen = generator.flow(valid_labels.index, valid_targets)
history = model.fit(
train_gen,
epochs=self.hyperparams["n_epochs"],
validation_data=valid_gen,
verbose=self.hyperparams["verbose"],
shuffle=True,
use_multiprocessing=True,
)
sg.utils.plot_history(history)
test_gen = generator.flow(test_labels.index, test_targets)
test_metrics = model.evaluate(test_gen)
print("\nTest Set Metrics:")
for name, valid in zip(model.metrics_names, test_metrics):
print("\t{}: {:0.4f}".format(name, valid))
def _dispatch_inductive_layer(layer_sizes, generator, model_name, params):
if model_name == "attri2vec":
embedding_layer = Attri2Vec(
layer_sizes=layer_sizes,
generator=generator,
bias=False, normalize=None
)
elif model_name == "graphsage":
embedding_layer = GraphSAGE(
layer_sizes=layer_sizes,
generator=generator,
bias=True,
dropout=0.0,
normalize="l2"
)
return embedding_layer
def create_graphSAGE_model(graph, link_prediction=False):
if link_prediction:
# We are going to train on the original graph
generator = GraphSAGELinkGenerator(graph,
batch_size=2,
num_samples=[2, 2])
edge_ids_train = np.array([[1, 2], [2, 3], [1, 3]])
train_gen = generator.flow(edge_ids_train, np.array([1, 1, 0]))
else:
generator = GraphSAGENodeGenerator(graph,
batch_size=2,
num_samples=[2, 2])
train_gen = generator.flow([1, 2], np.array([[1, 0], [0, 1]]))
# if link_prediction:
# edge_ids_train = np.array([[1, 2], [2, 3], [1, 3]])
# train_gen = generator.flow(edge_ids_train, np.array([1, 1, 0]))
# else:
# train_gen = generator.flow([1, 2], np.array([[1, 0], [0, 1]]))
base_model = GraphSAGE(layer_sizes=[8, 8],
generator=train_gen,
bias=True,
dropout=0.5)
if link_prediction:
# Expose input and output sockets of graphsage, for source and destination nodes:
x_inp_src, x_out_src = base_model.node_model()
x_inp_dst, x_out_dst = base_model.node_model()
# re-pack into a list where (source, destination) inputs alternate, for link inputs:
x_inp = [x for ab in zip(x_inp_src, x_inp_dst) for x in ab]
# same for outputs:
x_out = [x_out_src, x_out_dst]
prediction = link_classification(output_dim=1,
output_act="relu",
edge_embedding_method="ip")(x_out)
keras_model = Model(inputs=x_inp, outputs=prediction)
else:
x_inp, x_out = base_model.node_model()
prediction = layers.Dense(units=2, activation="softmax")(x_out)
keras_model = Model(inputs=x_inp, outputs=prediction)
return base_model, keras_model, generator, train_gen
def _fit_deep_graph_infomax(train_graph, params, model_name):
"""Train unsupervised Deep Graph Infomax."""
if "gcn_dgi" in model_name or "gat_dgi" in model_name:
if "cluster" in model_name:
generator = ClusterNodeGenerator(
train_graph, clusters=params["clusters"],
q=params["clusters_q"])
else:
generator = FullBatchNodeGenerator(train_graph, sparse=False)
if "gcn_dgi" in model_name:
embedding_layer = GCN(
layer_sizes=[params["embedding_dimension"]],
activations=["relu"], generator=generator)
elif "gat_dgi" in model_name:
embedding_layer = GAT(
layer_sizes=[params["embedding_dimension"]],
activations=["relu"], generator=generator, attn_heads=8)
elif model_name == "graphsage_dgi":
generator = GraphSAGENodeGenerator(
train_graph, batch_size=50, num_samples=[5])
embedding_layer = GraphSAGE(
layer_sizes=[params["embedding_dimension"]], activations=["relu"],
generator=generator
)
else:
raise ValueError(f"Unknown mode name {model_name}")
embedding_model = _execute_deep_graph_infomax(
train_graph, embedding_layer, generator, params)
# Here the models can be both inductive and transductive
if model_name in ["gcn_dgi", "gat_dgi", "graphsage_dgi"]:
return embedding_model.predict(
generator.flow(train_graph.nodes()))
else:
return embedding_model
layer_sizes = [40, 30]
learning_rate = 5e-2
unsupervisedSamples = UnsupervisedSampler(Gs,
nodes=G.nodes(),
length=length,
number_of_walks=number_of_walks)
generator = GraphSAGELinkGenerator(Gs, batch_size, num_samples)
train_gen = generator.flow(unsupervisedSamples)
assert len(layer_sizes) == len(num_samples)
graphsage = GraphSAGE(layer_sizes=layer_sizes,
generator=generator,
bias=True,
dropout=0.0,
normalize="l2")
x_inp, x_out = graphsage.build()
prediction = link_classification(output_dim=1,
output_act="sigmoid",
edge_embedding_method="ip")(x_out)
model = keras.Model(inputs=x_inp, outputs=prediction)
model.compile(optimizer=keras.optimizers.Adam(lr=learning_rate),
loss=keras.losses.binary_crossentropy,
metrics=[keras.metrics.binary_accuracy])
def train(
G,
layer_size: List[int],
num_samples: List[int],
batch_size: int = 100,
num_epochs: int = 10,
learning_rate: float = 0.001,
dropout: float = 0.0,
):
"""
Train the GraphSAGE model on the specified graph G
with given parameters.
Args:
G: NetworkX graph file
layer_size: A list of number of hidden units in each layer of the GraphSAGE model
num_samples: Number of neighbours to sample at each layer of the GraphSAGE model
batch_size: Size of batch for inference
num_epochs: Number of epochs to train the model
learning_rate: Initial Learning rate
dropout: The dropout (0->1)
"""
# Split links into train/test
print("Using '{}' method to sample negative links".format(
args.edge_sampling_method))
# From the original graph, extract E_test and the reduced graph G_test:
edge_splitter_test = EdgeSplitter(G)
# Randomly sample a fraction p=0.1 of all positive links, and same number of negative links, from G, and obtain the
# reduced graph G_test with the sampled links removed:
G_test, edge_ids_test, edge_labels_test = edge_splitter_test.train_test_split(
p=0.1,
keep_connected=True,
method=args.edge_sampling_method,
probs=args.edge_sampling_probs,
)
# From G_test, extract E_train and the reduced graph G_train:
edge_splitter_train = EdgeSplitter(G_test, G)
# Randomly sample a fraction p=0.1 of all positive links, and same number of negative links, from G_test, and obtain the
# further reduced graph G_train with the sampled links removed:
G_train, edge_ids_train, edge_labels_train = edge_splitter_train.train_test_split(
p=0.1,
keep_connected=True,
method=args.edge_sampling_method,
probs=args.edge_sampling_probs,
)
# G_train, edge_ds_train, edge_labels_train will be used for model training
# G_test, edge_ds_test, edge_labels_test will be used for model testing
# Convert G_train and G_test to StellarGraph objects (undirected, as required by GraphSAGE) for ML:
G_train = sg.StellarGraph(G_train, node_features="feature")
G_test = sg.StellarGraph(G_test, node_features="feature")
# Mapper feeds link data from sampled subgraphs to GraphSAGE model
# We need to create two mappers: for training and testing of the model
train_gen = GraphSAGELinkGenerator(G_train, batch_size, num_samples)
train_flow = train_gen.flow(edge_ids_train,
edge_labels_train,
shuffle=True)
test_gen = GraphSAGELinkGenerator(G_test, batch_size, num_samples)
test_flow = test_gen.flow(edge_ids_test, edge_labels_test)
# GraphSAGE model
graphsage = GraphSAGE(layer_sizes=layer_size,
generator=train_gen,
bias=True,
dropout=dropout)
# Construct input and output tensors for the link prediction model
x_inp, x_out = graphsage.build()
# Final estimator layer
prediction = link_classification(
output_dim=1,
output_act="sigmoid",
edge_embedding_method=args.edge_embedding_method,
)(x_out)
# Stack the GraphSAGE and prediction layers into a Keras model, and specify the loss
model = keras.Model(inputs=x_inp, outputs=prediction)
model.compile(
optimizer=optimizers.Adam(lr=learning_rate),
loss=losses.binary_crossentropy,
metrics=[metrics.binary_accuracy],
)
# Evaluate the initial (untrained) model on the train and test set:
init_train_metrics = model.evaluate_generator(train_flow)
init_test_metrics = model.evaluate_generator(test_flow)
print("\nTrain Set Metrics of the initial (untrained) model:")
for name, val in zip(model.metrics_names, init_train_metrics):
print("\t{}: {:0.4f}".format(name, val))
print("\nTest Set Metrics of the initial (untrained) model:")
for name, val in zip(model.metrics_names, init_test_metrics):
print("\t{}: {:0.4f}".format(name, val))
# Train model
print("\nTraining the model for {} epochs...".format(num_epochs))
history = model.fit_generator(
train_flow,
epochs=num_epochs,
validation_data=test_flow,
verbose=2,
shuffle=False,
)
# Evaluate and print metrics
train_metrics = model.evaluate_generator(train_flow)
test_metrics = model.evaluate_generator(test_flow)
print("\nTrain Set Metrics of the trained model:")
for name, val in zip(model.metrics_names, train_metrics):
print("\t{}: {:0.4f}".format(name, val))
print("\nTest Set Metrics of the trained model:")
for name, val in zip(model.metrics_names, test_metrics):
print("\t{}: {:0.4f}".format(name, val))
# Save the trained model
save_str = "_n{}_l{}_d{}_r{}".format(
"_".join([str(x) for x in num_samples]),
"_".join([str(x) for x in layer_size]),
dropout,
learning_rate,
)
model.save("graphsage_link_pred" + save_str + ".h5")
def train(
edgelist,
node_data,
layer_size,
num_samples,
batch_size=100,
num_epochs=10,
learning_rate=0.005,
dropout=0.0,
target_name="subject",
):
"""
Train a GraphSAGE model on the specified graph G with given parameters, evaluate it, and save the model.
Args:
edgelist: Graph edgelist
node_data: Feature and target data for nodes
layer_size: A list of number of hidden nodes in each layer
num_samples: Number of neighbours to sample at each layer
batch_size: Size of batch for inference
num_epochs: Number of epochs to train the model
learning_rate: Initial Learning rate
dropout: The dropout (0->1)
"""
# Extract target and encode as a one-hot vector
target_encoding = feature_extraction.DictVectorizer(sparse=False)
node_targets = target_encoding.fit_transform(
node_data[[target_name]].to_dict("records"))
node_ids = node_data.index
# Extract the feature data. These are the feature vectors that the Keras model will use as input.
# The CORA dataset contains attributes 'w_x' that correspond to words found in that publication.
node_features = node_data[feature_names]
# Create graph from edgelist and set node features and node type
Gnx = nx.from_pandas_edgelist(edgelist, edge_attr="label")
nx.set_node_attributes(Gnx, "paper", "label")
# Convert to StellarGraph and prepare for ML
G = sg.StellarGraph(Gnx,
node_type_name="label",
node_features=node_features)
# Split nodes into train/test using stratification.
train_nodes, test_nodes, train_targets, test_targets = model_selection.train_test_split(
node_ids,
node_targets,
train_size=140,
test_size=None,
stratify=node_targets,
random_state=5232,
)
# Split test set into test and validation
val_nodes, test_nodes, val_targets, test_targets = model_selection.train_test_split(
test_nodes,
test_targets,
train_size=500,
test_size=None,
random_state=5214)
# Create mappers for GraphSAGE that input data from the graph to the model
generator = GraphSAGENodeGenerator(G, batch_size, num_samples, seed=5312)
train_gen = generator.flow(train_nodes, train_targets, shuffle=True)
val_gen = generator.flow(val_nodes, val_targets)
# GraphSAGE model
model = GraphSAGE(
layer_sizes=layer_size,
generator=train_gen,
bias=True,
dropout=dropout,
aggregator=MeanAggregator,
)
# Expose the input and output sockets of the model:
x_inp, x_out = model.build()
# Snap the final estimator layer to x_out
prediction = layers.Dense(units=train_targets.shape[1],
activation="softmax")(x_out)
# Create Keras model for training
model = keras.Model(inputs=x_inp, outputs=prediction)
model.compile(
optimizer=optimizers.Adam(lr=learning_rate, decay=0.001),
loss=losses.categorical_crossentropy,
metrics=[metrics.categorical_accuracy],
)
print(model.summary())
# Train model
history = model.fit_generator(train_gen,
epochs=num_epochs,
validation_data=val_gen,
verbose=2,
shuffle=False)
# Evaluate on test set and print metrics
test_metrics = model.evaluate_generator(
generator.flow(test_nodes, test_targets))
print("\nTest Set Metrics:")
for name, val in zip(model.metrics_names, test_metrics):
print("\t{}: {:0.4f}".format(name, val))
# Get predictions for all nodes
all_predictions = model.predict_generator(generator.flow(node_ids))
# Turn predictions back into the original categories
node_predictions = pd.DataFrame(
target_encoding.inverse_transform(all_predictions), index=node_ids)
accuracy = np.mean([
"subject=" + gt_subject == p for gt_subject, p in zip(
node_data["subject"], node_predictions.idxmax(axis=1))
])
print("All-node accuracy: {:3f}".format(accuracy))
# TODO: extract the GraphSAGE embeddings from x_out, and save/plot them
# Save the trained model
save_str = "_n{}_l{}_d{}_r{}".format(
"_".join([str(x) for x in num_samples]),
"_".join([str(x) for x in layer_size]),
dropout,
learning_rate,
)
model.save("cora_example_model" + save_str + ".h5")
# We must also save the target encoding to convert model predictions
with open("cora_example_encoding" + save_str + ".pkl", "wb") as f:
pickle.dump([target_encoding], f)
train_ids = nodes[:5000]
test_ids = nodes[5000:]
train_labels= [graph.nodes[id]["_class"] for id in train_ids]
test_labels = [graph.nodes[id]["_class"] for id in test_ids]
all_labels = train_labels + test_labels
train_labels = np.array(train_labels).reshape(len(train_ids),1)
test_labels = np.array(test_labels).reshape(len(test_ids), 1)
print(np.unique(train_labels, return_counts=True))
print(np.unique(test_labels, return_counts=True))
generator = GraphSAGENodeGenerator(G, batch_size=50, num_samples=[10,10])
train_data_gen = generator.flow(train_ids, train_labels)
test_data_gen = generator.flow(test_ids, test_labels)
all_gen = generator.flow(list(nodes), all_labels)
print("Node Gen done!")
base_model = GraphSAGE(layer_sizes=[32, 32], generator=generator, bias=True, dropout=0.8)
x_in, x_out = base_model.build()
prediction = layers.Dense(units=2, activation="softmax")(x_out)
print("model building done")
model = Model(inputs=x_in, outputs = prediction)
model.compile(optimizer=optimizers.Adam(lr=0.005), loss=losses.categorical_crossentropy, metrics=["acc"])
tensorboard = callbacks.TensorBoard(log_dir="logs",embeddings_freq=1, update_freq=1, histogram_freq=1)
tboard = model.fit(train_data_gen, epochs=4, validation_data=test_data_gen, verbose=True,
shuffle=False, callbacks=[tensorboard])
print(tboard)
print("prediction done")
y_pred = model.predict(train_data_gen, verbose=1)
labels = np.argmax(y_pred, axis=1)
# together wi"number of parallel workers to use" the unsupervised sampler will be used to generate samples.
actual_nodes_train = list(Gtrain.nodes())
if testtype == 'nodes':
assert set(nodes_train).issuperset(actual_nodes_train)
unsupervised_samples = UnsupervisedSampler(Gtrain,
nodes=actual_nodes_train,
length=length_of_walks,
number_of_walks=number_of_walks)
train_gen = GraphSAGELinkGenerator(Gtrain, batch_size,
num_samples).flow(unsupervised_samples)
# Build the model
assert len(layer_sizes) == len(num_samples)
graphsage = GraphSAGE(layer_sizes=layer_sizes,
generator=train_gen,
bias=bias,
dropout=0.0,
normalize="l2")
x_inp, x_out = graphsage.build(flatten_output=False)
prediction = link_classification(output_dim=1,
output_act="sigmoid",
edge_embedding_method='ip')(x_out)
model = keras.Model(inputs=x_inp, outputs=prediction)
model.compile(
optimizer=keras.optimizers.Adam(lr=1e-3),
loss=keras.losses.binary_crossentropy,
metrics=[keras.metrics.binary_accuracy],
)
# Train the model
history = model.fit_generator(
# Hyperparams
batch_size = 20
epochs = 20
num_samples = [20, 10]
layer_sizes = [20, 20]
# Train iterators
train_gen = GraphSAGELinkGenerator(G_train, batch_size, num_samples)
train_flow = train_gen.flow(edge_ids_train, edge_labels_train, shuffle=True)
# Model defining - Keras functional API + Stellargraph layers
graphsage = GraphSAGE(
layer_sizes=layer_sizes, generator=train_gen, bias=True, dropout=0.3
)
x_inp, x_out = graphsage.in_out_tensors()
prediction = link_classification(
output_dim=1, output_act="relu", edge_embedding_method="ip"
)(x_out)
model = keras.Model(inputs=x_inp, outputs=prediction)
model.compile(
optimizer=keras.optimizers.Adam(lr=1e-3),
loss=keras.losses.binary_crossentropy,
metrics=["acc"],
)
def train(G_list,
nodes_subjects_list,
run_num=1,
start_month_id=220,
end_month_id=264):
# 提前定义一些列表方便记录数据，大循环的列表存小循环的列表
graph_history_list_list = []
model_list_list = []
train_gen_list_list = []
time_list_list = []
model_weight_list_list = []
# 选择运行run_num次
run_num = run_num
# 选择进行训练的月份,end_month_id最多取
start_month_id = start_month_id
end_month_id = end_month_id
# 创建文件夹保存model
if not os.path.exists('model'):
os.makedirs('model')
# 创建文件夹保存history
if not os.path.exists('history'):
os.makedirs('history')
# 创建文件夹保存figure
if not os.path.exists('figure'):
os.makedirs('figure')
# 创建文件夹保存figure
if not os.path.exists('figure_distribution'):
os.makedirs('figure_distribution')
# 创建文件夹保存test结果
if not os.path.exists('test_result'):
os.makedirs('test_result')
# 大循环记录训练了几次，计算多次是为了减少variance
# 小循环记录训练的月份
for j in range(run_num):
num_samples = [40]
# 提前定义一些列表记录小循环的数据
graph_history_list = []
model_list = []
train_gen_list = []
time_list = []
model_weight_list = []
test_result = []
# i为0代表220
for i in range(start_month_id - 220, end_month_id - 220):
start = time.time()
# 前一个月训练，后一个月验证
train_idx = i
val_idx = i + 1
test_idx = i + 2
# 用train_idx的数据生成训练集的generator
generator = GraphSAGENodeGenerator(
G=G_list[train_idx],
batch_size=len(nodes_subjects_list[train_idx]),
num_samples=num_samples,
seed=100)
train_gen = generator.flow(list(
nodes_subjects_list[train_idx].index),
nodes_subjects_list[train_idx].values,
shuffle=False)
# 生成GraphSAGE模型
graphsage_model = GraphSAGE(layer_sizes=[1],
generator=generator,
bias=True,
aggregator=sg.layer.MeanAggregator,
normalize=None)
# 提取输出输出的tensor，用keras来构建模型
x_inp, x_out = graphsage_model.in_out_tensors()
# prediction = layers.Dense(units=1)(x_out)
# 用val_idx的数据生成验证集的generator
generator = GraphSAGENodeGenerator(
G=G_list[val_idx],
batch_size=len(nodes_subjects_list[val_idx]),
num_samples=num_samples,
seed=100)
val_gen = generator.flow(list(nodes_subjects_list[val_idx].index),
nodes_subjects_list[val_idx].values)
# 用test_idx的数据生成验证集的generator
generator = GraphSAGENodeGenerator(
G=G_list[test_idx],
batch_size=len(nodes_subjects_list[test_idx]),
num_samples=num_samples,
seed=100)
test_gen = generator.flow(
list(nodes_subjects_list[test_idx].index),
nodes_subjects_list[test_idx].values)
# 通过输入输出的tensor构建model
model = Model(inputs=x_inp, outputs=x_out)
monitor = EarlyStopping(monitor='val_loss',
min_delta=1e-3,
patience=10,
verbose=2,
mode='auto',
restore_best_weights=True)
model.compile(optimizer=optimizers.Adam(lr=0.05),
loss=losses.mean_squared_error,
metrics=[pearson_r])
history = model.fit(train_gen,
epochs=500,
validation_data=val_gen,
verbose=0,
shuffle=False,
callbacks=[monitor])
test_metrics = model.evaluate(test_gen)
test_result_dict = {}
print("\n" + str(train_idx + 220) + "'s Test Set: " +
str(test_idx + 220) + "'s Metrics:")
for name, val in zip(model.metrics_names, test_metrics):
print("\t{}: {:0.4f}".format(name, val))
test_result_dict[name] = val
json.dump(
test_result_dict,
open(
'test_result/' + str(train_idx + 220) + "_" +
str(test_idx + 220) + '.json', 'w'))
test_preds = model.predict(test_gen)
end = time.time()
# 保存一些结果
graph_history_list.append(history) # 保存训练过程
model_list.append(model) # 保存model
train_gen_list.append(train_gen) # 保存train_gen方便之后算中间层的结果
time_list.append(end - start) # 保存运行时间
model_weight_list.append(model.weights) # 保存model的参数
test_result.append(test_metrics[1])
# # 存模型model
# model.save('model/' + str(train_idx + 220) + "_" + str(val_idx + 220) + '.h5')
# # 存训练过程history
# json.dump(history.history,
# open('history/' + str(train_idx + 220) + "_" + str(val_idx + 220) + '.json', 'w'))
# # 存训练过程图片figure
# sg.utils.plot_history(history)
# plt.title(str(train_idx + 220) + '->' + str(val_idx + 220))
# plt.savefig('figure/' + str(train_idx + 220) + "_" + str(val_idx + 220) + '.png')
# plt.show()
# 存test的prediction的distribution
plt.figure(figsize=(5, 10))
plt.subplot(211)
plt.hist(test_preds, bins=500)
plt.title("Distribution of Prediction of " + str(test_idx + 220))
plt.subplot(212)
plt.hist(nodes_subjects_list[test_idx].values, bins=500)
plt.title("Distribution of Origin of " + str(test_idx + 220))
plt.xlabel("ic=" + str(test_metrics[1]))
plt.savefig('figure_distribution/distribution-' +
str(train_idx + 220) + "_" + str(test_idx + 220) +
'.png',
dpi=300)
plt.show()
print(str(i + 220) + "'s " + str(j + 1) + " run has finished")
print()
# 将小循环的数据保存
graph_history_list_list.append(graph_history_list)
model_list_list.append(model_list)
train_gen_list_list.append(train_gen_list)
time_list_list.append(time_list)
model_weight_list_list.append(model_weight_list)
return graph_history_list_list, model_list_list, train_gen_list_list, time_list_list, model_weight_list_list, test_result
def train_model(Gnx, train_data, test_data, all_features):
output_results = {}
from collections import Counter
#TODO: save size of dataset, train_data, and test data
#save the count of each subject in the blocks
print(len(train_data), len(test_data))
subject_groups_train = Counter(train_data['subject'])
subject_groups_test = Counter(test_data['subject'])
output_results['train_size'] = len(train_data)
output_results['test_size'] = len(test_data)
output_results['subject_groups_train'] = subject_groups_train
output_results['subject_groups_test'] = subject_groups_test
#node_features = train_data[feature_names]
#print (feature_names)
G = sg.StellarGraph(Gnx, node_features=all_features)
#TODO: save graph info
print(G.info())
print("writing graph.dot")
#write_dot(Gnx,"graph.dot")
output_results['graph_info'] = G.info()
print("building the graph generator...")
batch_size = 50
num_samples = [10, 5]
generator = GraphSAGENodeGenerator(G, batch_size, num_samples)
#generator = HinSAGENodeGenerator(G, batch_size, num_samples)
target_encoding = feature_extraction.DictVectorizer(sparse=False)
train_targets = target_encoding.fit_transform(
train_data[["subject"]].to_dict('records'))
print(np.unique(train_data["subject"].to_list()))
class_weights = class_weight.compute_class_weight(
'balanced', np.unique(train_data["subject"].to_list()),
train_data["subject"].to_list())
print('class_weights', class_weights)
test_targets = target_encoding.transform(test_data[["subject"
]].to_dict('records'))
train_gen = generator.flow(train_data.index, train_targets, shuffle=True)
graphsage_model = GraphSAGE(
#graphsage_model = HinSAGE(
#layer_sizes=[32, 32],
layer_sizes=[80, 80],
generator=generator, #train_gen,
bias=True,
dropout=0.5,
)
print("building model...")
#x_inp, x_out = graphsage_model.build(flatten_output=True)
x_inp, x_out = graphsage_model.build()
prediction = layers.Dense(units=train_targets.shape[1],
activation="softmax")(x_out)
model = Model(inputs=x_inp, outputs=prediction)
print("compiling model...")
model.compile(
optimizer=optimizers.Adam(lr=0.005),
loss=losses.categorical_crossentropy,
metrics=["acc", metrics.categorical_accuracy],
)
print("testing the model...")
test_gen = generator.flow(test_data.index, test_targets)
history = model.fit_generator(
train_gen,
epochs=EPOCH,
validation_data=test_gen,
verbose=2,
shuffle=True,
class_weight=class_weights,
)
# save test metrics
test_metrics = model.evaluate_generator(test_gen)
print("\nTest Set Metrics:")
output_results['test_metrics'] = []
for name, val in zip(model.metrics_names, test_metrics):
output_results['test_metrics'].append({'name': name, 'val:': val})
print("\t{}: {:0.4f}".format(name, val))
test_nodes = test_data.index
test_mapper = generator.flow(test_nodes)
test_predictions = model.predict_generator(test_mapper)
node_predictions = target_encoding.inverse_transform(test_predictions)
results = pd.DataFrame(node_predictions, index=test_nodes).idxmax(axis=1)
df = pd.DataFrame({
"Predicted": results,
"True": test_data['subject']
}) #, "program":test_data['program']})
clean_result_labels = df["Predicted"].map(
lambda x: x.replace('subject=', ''))
# save predicted labels
pred_labels = np.unique(clean_result_labels.values)
#pred_program = np.unique(df['program'].values)
# save predictions per label
precision, recall, f1, _ = skmetrics.precision_recall_fscore_support(
df['True'].values,
clean_result_labels.values,
average=None,
labels=pred_labels)
output_results['classifier'] = []
for lbl, prec, rec, fm in zip(pred_labels, precision, recall, f1):
output_results['classifier'].append({
'label': lbl,
'precision': prec,
'recall': rec,
'fscore': fm
})
print(output_results['classifier'])
print(pred_labels)
print('precision: {}'.format(precision))
print('recall: {}'.format(recall))
print('fscore: {}'.format(f1))
return generator, model, x_inp, x_out, history, target_encoding, output_results
# on concatenated `(paper1, paper2)` node embeddings.
#
# GraphSAGE part of the model, with hidden layer sizes of 20 for both GraphSAGE layers, a bias term, and no dropout.
# (Dropout can be switched on by specifying a positive dropout rate, 0 < dropout < 1)
#
# Note that the length of layer_sizes list must be equal to the length of num_samples, as len(num_samples) defines
# the number of hops (layers) in the GraphSAGE model.
# In[17]:
layer_sizes = [20, 20]
assert len(layer_sizes) == len(num_samples)
graphsage = GraphSAGE(
layer_sizes=layer_sizes, generator=generator, bias=True, dropout=0.5
)
# In[18]:
# Build the model and expose the input and output tensors.
x_inp, x_out = graphsage.build()
# Final link classification layer that takes a pair of node embeddings produced by graphsage, applies a binary
# operator to them to produce the corresponding link embedding ('ip' for inner product; other options for the binary
# operator can be seen by running a cell with `?link_classification` in it), and passes it through a dense layer:
# In[19]:
# Create data generator for our graph, specified by which type
# of model (GraphSAGE) and the learning task (Node) ...
generator = GraphSAGENodeGenerator(G, batch_size, num_samples)
print(train_subjects.index)
# Create an iterator for our training data, this takes the indeces of the
# nodes in the graph to be used for training, as well as their respective
# one-hot encoded label vectors
train_gen = generator.flow(train_subjects.index, train_targets, shuffle = True)
# Specify the graph-learning model
graphsage_model = GraphSAGE(
layer_sizes=[32, 32], generator=generator, bias=True, dropout=0.5,
aggregator = MeanAggregator
)
# Extract the input and output tensors of the model. Set predictions
# of the model to be a softmax layer taking output tensor as its input.
x_inp, x_out = graphsage_model.in_out_tensors()
prediction = layers.Dense(units=train_targets.shape[1], activation="softmax")(x_out)
model = Model(inputs = x_inp, outputs = prediction)
model.compile(
optimizer=optimizers.Adam(lr=0.005),
loss=losses.categorical_crossentropy,
metrics=["acc"],
)
# To validate/test we need another generator for testing data, no shuffle needed
test_gen = generator.flow(test_subjects.index, test_targets)
history = model.fit(
test_gen = generator.flow(test_subjects.index, test_targets)
# aggregatortype = MaxPoolingAggregator(),
# layer_sizes (list): Hidden feature dimensions for each layer. activations (list): Activations applied to each layer's output;
def get_dropout(input_tensor, p=0.1, mc=False):
if mc:
return Dropout(p)(input_tensor, training=True)
else:
return Dropout(p)(input_tensor)
graphsage_model = GraphSAGE(layer_sizes=[64, 32, 16],
generator=generator,
activations=["relu", "relu", "linear"],
bias=True,
aggregator=MaxPoolingAggregator,
dropout=0.1)
x_inp, x_out = graphsage_model.in_out_tensors()
x_out = layers.Dense(units=10, activation="relu")(x_out)
x_out = layers.Dense(units=10, activation="relu")(x_out)
x_out = get_dropout(x_out, p=0.1, mc='mc')
prediction = layers.Dense(units=train_targets.shape[1],
activation="softmax")(x_out)
model = Model(inputs=x_inp, outputs=prediction)
model.summary()
##
# model.compile( optimizer=optimizers.Adam(), loss = noderankloss(), metrics=["acc"])
def _train_model(self, gnx, train_data, test_data, all_features,
target_feature_name):
subject_groups_train = Counter(train_data[target_feature_name])
subject_groups_test = Counter(test_data[target_feature_name])
graph = sg.StellarGraph(gnx, node_features=all_features)
output_results = {
'train_size': len(train_data),
'test_size': len(test_data),
'subject_groups_train': subject_groups_train,
'subject_groups_test': subject_groups_test,
'graph_info': graph.info()
}
num_samples = [10, 5]
generator = GraphSAGENodeGenerator(graph, self.batch_size, num_samples)
target_encoding = feature_extraction.DictVectorizer(sparse=False)
train_targets = target_encoding.fit_transform(
train_data[[target_feature_name]].to_dict('records'))
class_weights = class_weight.compute_class_weight(
class_weight='balanced',
classes=np.unique(train_data[target_feature_name].to_list()),
y=train_data[target_feature_name].to_list())
class_weights = dict(enumerate(class_weights))
test_targets = target_encoding.transform(
test_data[[target_feature_name]].to_dict('records'))
train_gen = generator.flow(train_data.index,
train_targets,
shuffle=True)
graph_sage_model = GraphSAGE(
layer_sizes=[80, 80],
generator=generator, # train_gen,
bias=True,
dropout=0.5,
)
print('building model...')
x_inp, x_out = graph_sage_model.build()
prediction = layers.Dense(units=train_targets.shape[1],
activation="softmax")(x_out)
model = Model(inputs=x_inp, outputs=prediction)
print('compiling model...')
model.compile(
optimizer=optimizers.Adam(learning_rate=0.005),
loss=losses.categorical_crossentropy,
metrics=['acc', metrics.categorical_accuracy],
)
print('testing the model...')
test_gen = generator.flow(test_data.index, test_targets)
history = model.fit(
train_gen,
epochs=self.num_epochs,
validation_data=test_gen,
verbose=2,
shuffle=True,
class_weight=class_weights,
)
# save test metrics
test_metrics = model.evaluate(test_gen)
print('Test Set Metrics:')
output_results['test_metrics'] = []
for name, val in zip(model.metrics_names, test_metrics):
output_results['test_metrics'].append({'name': name, 'val:': val})
print("\t{}: {:0.4f}".format(name, val))
test_nodes = test_data.index
test_mapper = generator.flow(test_nodes)
test_predictions = model.predict(test_mapper)
node_predictions = target_encoding.inverse_transform(test_predictions)
results = pd.DataFrame(node_predictions,
index=test_nodes).idxmax(axis=1)
df = pd.DataFrame({
'Predicted': results,
'True': test_data[target_feature_name]
})
clean_result_labels = df['Predicted'].map(
lambda x: x.replace('subject=', ''))
# save predicted labels
pred_labels = np.unique(clean_result_labels.values)
precision, recall, f1, _ = skmetrics.precision_recall_fscore_support(
df['True'].values,
clean_result_labels.values,
average=None,
labels=pred_labels)
output_results['classifier'] = []
for lbl, prec, rec, fm in zip(pred_labels, precision, recall, f1):
output_results['classifier'].append({
'label': lbl,
'precision': prec,
'recall': rec,
'fscore': fm
})
print(output_results['classifier'])
print(pred_labels)
print('precision: {}'.format(precision))
print('recall: {}'.format(recall))
print('fscore: {}'.format(f1))
output_results['history'] = {
'epochs': history.epoch,
'training_log': history.history,
'training_params': history.params
}
return generator, model, x_inp, x_out, history, target_encoding, output_results