def _build_model(self, graph: Graph) -> Model:
"""Return SkipGram model."""
# Create first the input with the central terms
central_terms = Input((1, ), dtype=tf.int32)
# Then we create the input of the contextual terms
contextual_terms = Input((self._window_size * 2, ), dtype=tf.int32)
# Creating the embedding layer for the contexts
central_term_embedding = Flatten()(Embedding(
input_dim=graph.get_number_of_nodes(),
output_dim=self._embedding_size,
input_length=1,
name=self.NODE_EMBEDDING,
)(central_terms))
# Adding layer that also executes the loss function
output = NoiseContrastiveEstimation(
vocabulary_size=graph.get_number_of_nodes(),
embedding_size=self._embedding_size,
number_of_negative_samples=self._number_of_negative_samples,
positive_samples=self._window_size * 2,
)((central_term_embedding, contextual_terms))
# Creating the actual model
model = Model(inputs=[contextual_terms, central_terms],
outputs=output,
name=self.model_name())
model.compile(optimizer=self._optimizer)
return model
def _fit_transform(
self,
graph: Graph,
return_dataframe: bool = True,
verbose: bool = True
) -> EmbeddingResult:
"""Return node embedding."""
edges, weights = graph.get_symmetric_normalized_laplacian_coo_matrix()
coo = coo_matrix(
(weights, (edges[:, 0], edges[:, 1])),
shape=(
graph.get_number_of_nodes(),
graph.get_number_of_nodes()
),
dtype=np.float32
)
embedding = eigsh(
coo,
k=self._embedding_size + 1,
which="LM",
return_eigenvectors=True
)[1]
if return_dataframe:
node_names = graph.get_node_names()
embedding = pd.DataFrame(
embedding,
index=node_names
)
return EmbeddingResult(
embedding_method_name=self.model_name(),
node_embeddings=embedding
)
def _fit(
self,
graph: Graph,
support: Optional[Graph] = None,
node_features: Optional[List[np.ndarray]] = None,
node_type_features: Optional[List[np.ndarray]] = None,
edge_features: Optional[List[np.ndarray]] = None,
):
"""Run fitting on the provided graph.
Parameters
--------------------
graph: Graph
The graph to run predictions on.
support: Optional[Graph] = None
The graph describiding the topological structure that
includes also the above graph. This parameter
is mostly useful for topological classifiers
such as Graph Convolutional Networks.
node_features: Optional[List[np.ndarray]] = None
The node features to use.
node_type_features: Optional[List[np.ndarray]] = None
The node type features to use.
edge_features: Optional[List[np.ndarray]] = None
The edge features to use.
"""
lpt = EdgePredictionTransformer(method=self._edge_embedding_method,
aligned_node_mapping=True)
lpt.fit(node_features)
if support is None:
support = graph
negative_graph = graph.sample_negative_graph(
number_of_negative_samples=int(
math.ceil(graph.get_number_of_edges() *
self._training_unbalance_rate)),
random_state=self._random_state,
sample_only_edges_with_heterogeneous_node_types=self.
_training_sample_only_edges_with_heterogeneous_node_types,
use_zipfian_sampling=self._use_zipfian_sampling)
if self._use_edge_metrics:
self._support = support
edge_features = np.vstack((support.get_all_edge_metrics(
normalize=True,
subgraph=graph,
),
support.get_all_edge_metrics(
normalize=True,
subgraph=negative_graph,
)))
self._model_instance.fit(
*lpt.transform(positive_graph=graph,
negative_graph=negative_graph,
edge_features=edge_features,
shuffle=True,
random_state=self._random_state))
def _build_model(self, graph: Graph) -> Model:
"""Return CBOW model."""
# Creating the inputs layers
# Create first the input with the central terms
central_terms = Input((1, ), dtype=tf.int32)
# Then we create the input of the contextual terms
contextual_terms = Input((self._window_size * 2, ), dtype=tf.int32)
# Getting the average context embedding
average_context_embedding = GlobalAveragePooling1D()(Embedding(
input_dim=graph.get_number_of_nodes(),
output_dim=self._embedding_size,
input_length=self._window_size * 2,
name="node_embedding",
)(contextual_terms))
# Adding layer that also executes the loss function
sampled_softmax = SampledSoftmax(
vocabulary_size=graph.get_number_of_nodes(),
embedding_size=self._embedding_size,
number_of_negative_samples=self._number_of_negative_samples,
)((average_context_embedding, central_terms))
# Creating the actual model
model = Model(inputs=[contextual_terms, central_terms],
outputs=sampled_softmax,
name=self.model_name())
model.compile(optimizer=self._optimizer)
return model
def __init__(self,
graph: Graph,
use_node_types: bool = False,
use_edge_metrics: bool = False,
batch_size: int = 2**10,
negative_samples_rate: float = 0.5,
avoid_false_negatives: bool = False,
graph_to_avoid: Graph = None,
sample_only_edges_with_heterogeneous_node_types: bool = False,
random_state: int = 42):
"""Create new EdgePredictionSequence object.
Parameters
--------------------------------
graph: Graph,
The graph from which to sample the edges.
use_node_types: bool = False,
Whether to return the node types.
use_edge_metrics: bool = False,
Whether to return the edge metrics.
batch_size: int = 2**10,
The batch size to use.
negative_samples_rate: float = 0.5,
Factor of negatives to use in every batch.
For example, with a batch size of 128 and negative_samples_rate equal
to 0.5, there will be 64 positives and 64 negatives.
avoid_false_negatives: bool = False,
Whether to filter out false negatives.
By default False.
Enabling this will slow down the batch generation while (likely) not
introducing any significant gain to the model performance.
graph_to_avoid: Graph = None,
Graph to avoid when generating the edges.
This can be the validation component of the graph, for example.
More information to how to generate the holdouts is available
in the Graph package.
sample_only_edges_with_heterogeneous_node_types: bool = False
Whether to only sample edges between heterogeneous node types.
This may be useful when training a model to predict between
two portions in a bipartite graph.
random_state: int = 42,
The random_state to use to make extraction reproducible.
"""
if not graph.has_edges():
raise ValueError(
f"An empty instance of graph {graph.get_name()} was provided!")
self._graph = graph
self._negative_samples_rate = negative_samples_rate
self._avoid_false_negatives = avoid_false_negatives
self._graph_to_avoid = graph_to_avoid
self._random_state = random_state
self._use_node_types = use_node_types
self._use_edge_metrics = use_edge_metrics
self._sample_only_edges_with_heterogeneous_node_types = sample_only_edges_with_heterogeneous_node_types
self._current_index = 0
super().__init__(
sample_number=graph.get_number_of_directed_edges(),
batch_size=batch_size,
)
def _get_class_weights(self, graph: Graph) -> Dict[int, float]:
"""Returns dictionary with class weights."""
nodes_number = graph.get_number_of_nodes()
node_types_number = graph.get_number_of_node_types()
return {
node_type_id: nodes_number / count / node_types_number
for node_type_id, count in graph.get_node_type_id_counts_hashmap().items()
}
def _get_class_weights(self, graph: Graph) -> Dict[int, float]:
"""Returns dictionary with class weights."""
number_of_directed_edges = graph.get_number_of_directed_edges()
edge_types_number = graph.get_number_of_edge_types()
return {
edge_type_id: number_of_directed_edges / count / edge_types_number
for edge_type_id, count in
graph.get_edge_type_id_counts_hashmap().items()
}
def _get_model_training_output(
self,
graph: Graph,
) -> Optional[np.ndarray]:
"""Returns training output tuple."""
if self.is_multilabel_prediction_task():
return graph.get_one_hot_encoded_node_types()
if self.is_binary_prediction_task():
return graph.get_boolean_node_type_ids()
return graph.get_single_label_node_type_ids()
def _extract_embeddings(self, graph: Graph,
model: Union[EntityRelationEmbeddingModel,
ERModel],
return_dataframe: bool) -> EmbeddingResult:
"""Returns embedding from the model.
Parameters
------------------
graph: Graph
The graph that was embedded.
model: Type[Model]
The Keras model used to embed the graph.
return_dataframe: bool
Whether to return a dataframe of a numpy array.
"""
if isinstance(model, EntityRelationEmbeddingModel):
node_embeddings = [model.entity_embeddings]
edge_type_embeddings = [model.relation_embeddings]
elif isinstance(model, ERModel):
node_embeddings = model.entity_representations
edge_type_embeddings = model.relation_representations
else:
raise NotImplementedError(
f"The provided model has type {type(model)}, which "
"is not currently supported. The supported types "
"are `EntityRelationEmbeddingModel` and `ERModel`.")
node_embeddings = [
node_embedding._embeddings.weight.cpu().detach().numpy()
for node_embedding in node_embeddings
]
edge_type_embeddings = [
edge_type_embedding._embeddings.weight.cpu().detach().numpy()
for edge_type_embedding in edge_type_embeddings
]
if return_dataframe:
node_embeddings = [
pd.DataFrame(node_embedding, index=graph.get_node_names())
for node_embedding in node_embeddings
]
edge_type_embeddings = [
pd.DataFrame(edge_type_embedding,
index=graph.get_unique_edge_type_names())
for edge_type_embedding in edge_type_embeddings
]
return EmbeddingResult(embedding_method_name=self.model_name(),
node_embeddings=node_embeddings,
edge_type_embeddings=edge_type_embeddings)
def _fit_transform(self,
graph: Graph,
return_dataframe: bool = True,
verbose: bool = True) -> EmbeddingResult:
"""Return node embedding."""
number_of_nodes = graph.get_number_of_nodes()
embedding = eigh(graph.get_dense_modularity_matrix(),
eigvals=(number_of_nodes - self._embedding_size,
number_of_nodes - 1))[1]
if return_dataframe:
node_names = graph.get_node_names()
embedding = pd.DataFrame(embedding, index=node_names)
return EmbeddingResult(embedding_method_name=self.model_name(),
node_embeddings=embedding)
def _get_model_training_input(
self,
graph: Graph,
support: Graph,
node_features: Optional[List[np.ndarray]] = None,
node_type_features: Optional[List[np.ndarray]] = None,
edge_features: Optional[List[np.ndarray]] = None,
) -> Tuple[Union[np.ndarray, Type[Sequence]]]:
"""Returns training input tuple."""
kernel = self.convert_graph_to_kernel(support)
return (
*(
()
if kernel is None
else (kernel,)
),
*(
()
if node_features is None
else node_features
),
*(
(graph.get_node_ids(),)
if self._use_node_embedding
else ()
)
)
def __init__(
self,
graph: Graph,
graph_used_in_training: Graph,
use_node_types: bool,
use_edge_metrics: bool,
batch_size: int = 2**10,
):
"""Create new EdgePredictionSequence object.
Parameters
--------------------------------
graph: Graph
The graph whose edges are to be predicted.
graph_used_in_training: Graph
The graph that was used while training the current
edge prediction model.
use_node_types: bool
Whether to return the node types.
use_edge_metrics: bool = False
Whether to return the edge metrics.
batch_size: int = 2**10,
The batch size to use.
"""
self._sequence = GenericEdgePredictionSequence(
graph=graph,
graph_used_in_training=graph_used_in_training,
use_node_types=use_node_types,
use_edge_metrics=use_edge_metrics,
batch_size=batch_size)
self._current_index = 0
super().__init__(
sample_number=graph.get_number_of_directed_edges(),
batch_size=batch_size,
)
def _extract_embeddings(self, graph: Graph, model: Model,
return_dataframe: bool) -> EmbeddingResult:
"""Returns embedding from the model.
Parameters
------------------
graph: Graph
The graph that was embedded.
model: Model
The Keras model used to embed the graph.
return_dataframe: bool
Whether to return a dataframe of a numpy array.
"""
node_embeddings = self.get_layer_weights(
"node_embeddings",
model,
)
context_embeddings = self.get_layer_weights(
"context_embeddings",
model,
)
if return_dataframe:
node_names = graph.get_node_names()
node_embeddings = pd.DataFrame(node_embeddings, index=node_names)
context_embeddings = pd.DataFrame(context_embeddings,
index=node_names)
return EmbeddingResult(
embedding_method_name=self.model_name(),
node_embeddings=[node_embeddings, context_embeddings])
def load_graph(self) -> Graph:
"""
Loads graph nodes and edges into Ensmallen.
Creates a node type list, as Ensmallen
requires this to parse node types.
:param graph_args: dict, output of main_graph_args
:return: ensmallen Graph
"""
graph_args_with_indir = self.main_graph_args()
for pathtype in ['node_path', 'edge_path']:
filepath = graph_args_with_indir[pathtype]
if is_url(filepath):
url_as_filename = \
''.join(c if c in VALID_CHARS else "_" for c in filepath)
outfile = os.path.join(self.outdir(), url_as_filename)
download_file(filepath, outfile)
graph_args_with_indir[pathtype] = outfile
elif not is_valid_path(filepath):
raise FileNotFoundError(f"Please check path: {filepath}")
# Now load the Ensmallen graph
loaded_graph = Graph.from_csv(**graph_args_with_indir)
return loaded_graph
def fit(
self,
graph: Graph,
support: Optional[Graph] = None,
node_features: Optional[Union[pd.DataFrame, np.ndarray,
List[Union[pd.DataFrame,
np.ndarray]]]] = None,
node_type_features: Optional[Union[pd.DataFrame, np.ndarray,
List[Union[pd.DataFrame,
np.ndarray]]]] = None,
edge_features: Optional[Union[pd.DataFrame, np.ndarray,
List[Union[pd.DataFrame,
np.ndarray]]]] = None,
):
"""Execute predictions on the provided graph.
Parameters
--------------------
graph: Graph
The graph to run predictions on.
support: Optional[Graph] = None
The graph describiding the topological structure that
includes also the above graph. This parameter
is mostly useful for topological classifiers
such as Graph Convolutional Networks.
node_features: Optional[Union[pd.DataFrame, np.ndarray, List[Union[pd.DataFrame, np.ndarray]]]] = None
The node features to use.
node_type_features: Optional[Union[pd.DataFrame, np.ndarray, List[Union[pd.DataFrame, np.ndarray]]]] = None
The node type features to use.
edge_features: Optional[Union[pd.DataFrame, np.ndarray, List[Union[pd.DataFrame, np.ndarray]]]] = None
The edge features to use.
"""
if node_type_features is not None:
raise NotImplementedError(
"Support for node type features is not currently available for any "
"of the edge-label prediction models.")
self._is_binary_prediction_task = graph.get_number_of_edge_types() == 2
self._is_multilabel_prediction_task = graph.is_multigraph()
super().fit(
graph=graph,
support=support,
node_features=node_features,
node_type_features=node_type_features,
edge_features=edge_features,
)
def convert_ensmallen_graph_to_networkx_graph(graph: Graph) -> nx.Graph:
"""Return networkX graph derived from the provided Ensmallen Graph.
Parameters
-----------
graph: Graph
The graph to be converted.
"""
if graph.is_directed():
result_graph = nx.DiGraph()
else:
result_graph = nx.Graph()
result_graph.add_nodes_from(graph.get_node_ids())
if graph.has_edge_weights():
result_graph.add_weighted_edges_from([
(src_name, dst_name, edge_weight)
for (src_name, dst_name), edge_weight in zip(
graph.get_directed_edge_node_ids(), graph.get_edge_weights())
])
else:
result_graph.add_edges_from(
graph.get_edge_node_ids(directed=graph.is_directed()))
return result_graph
def _get_steps_per_epoch(self, graph: Graph) -> int:
"""Returns number of steps per epoch.
Parameters
------------------
graph: Graph
The graph to compute the number of steps.
"""
return max(graph.get_number_of_directed_edges() // self._batch_size, 1)
def _get_steps_per_epoch(self, graph: Graph) -> Tuple[Any]:
"""Returns number of steps per epoch.
Parameters
------------------
graph: Graph
The graph to compute the number of steps.
"""
return max(graph.get_number_of_nodes() // self._batch_size, 1)
def __init__(
self,
graph: Graph,
graph_used_in_training: Graph,
return_node_types: bool,
return_edge_types: bool,
use_edge_metrics: bool,
batch_size: int = 2**10,
):
"""Create new EdgePredictionSequence object.
Parameters
--------------------------------
graph: Graph
The graph whose edges are to be predicted.
graph_used_in_training: Graph
The graph that was used while training the current
edge prediction model.
return_node_types: bool
Whether to return the node types.
return_edge_types: bool
Whether to return the edge types.
use_edge_metrics: bool = False
Whether to return the edge metrics.
batch_size: int = 2**10,
The batch size to use.
"""
if not graph.has_edges():
raise ValueError(
f"An empty instance of graph {graph.get_name()} was provided!")
if not graph.has_edges():
raise ValueError(
f"An empty instance of graph {graph_used_in_training.get_name()} was provided!"
)
if not graph.has_compatible_node_vocabularies(graph_used_in_training):
raise ValueError(
f"The provided graph {graph.get_name()} does not have a node vocabulary "
"that is compatible with the provided graph used in training.")
self._graph = graph
self._graph_used_in_training = graph_used_in_training
self._return_node_types = return_node_types
self._return_edge_types = return_edge_types
self._use_edge_metrics = use_edge_metrics
self._batch_size = batch_size
def split_graph_following_evaluation_schema(
cls,
graph: Graph,
evaluation_schema: str,
random_state: int,
holdout_number: int,
number_of_holdouts: int,
**holdouts_kwargs: Dict[str, Any],
) -> Tuple[Graph]:
"""Return train and test graphs tuple following the provided evaluation schema.
Parameters
----------------------
graph: Graph
The graph to split.
evaluation_schema: str
The evaluation schema to follow.
random_state: int
The random state for the evaluation
holdout_number: int
The current holdout number.
number_of_holdouts: int
The number of holdouts that will be generated throught the evaluation.
holdouts_kwargs: Dict[str, Any]
The kwargs to be forwarded to the holdout method.
"""
if evaluation_schema == "Stratified Monte Carlo":
return graph.get_edge_label_holdout_graphs(
**holdouts_kwargs,
use_stratification=True,
random_state=random_state + holdout_number,
)
if evaluation_schema == "Stratified Kfold":
return graph.get_edge_label_kfold(
k=number_of_holdouts,
k_index=holdout_number,
use_stratification=True,
random_state=random_state,
)
raise ValueError(
f"The requested evaluation schema `{evaluation_schema}` "
"is not available. The available evaluation schemas "
f"are: {format_list(cls.get_available_evaluation_schemas())}.")
def make_link_prediction_data(self, embedding_file: str,
training_graph_args: dict,
pos_validation_args: dict,
neg_training_args: dict,
neg_validation_args: dict,
edge_method: str) -> Tuple[Tuple, Tuple]:
"""Prepare training and validation data for training link prediction classifers
Args:
embedding_file: path to embedding file for nodes in graph
training_graph_args: EnsmallenGraph arguments to load training graph
pos_validation_args: EnsmallenGraph arguments to load positive validation graph
neg_training_args: EnsmallenGraph arguments to load negative training graph
neg_validation_args: EnsmallenGraph arguments to load negative validation graph
edge_method: edge embedding method to use (average, L1, L2, etc)
Returns:
A tuple of tuples
"""
embedding = pd.read_csv(embedding_file, index_col=0, header=None)
# load graphs
graphs = {'pos_training': Graph.from_csv(**training_graph_args)}
for name, graph_args in [('pos_validation', pos_validation_args),
('neg_training', neg_training_args),
('neg_validation', neg_validation_args)]:
these_params = copy.deepcopy(training_graph_args)
these_params.update(graph_args)
graphs[name] = Graph.from_csv(**these_params)
# create transformer object to convert graphs into edge embeddings
lpt = LinkPredictionTransformer(method=edge_method)
lpt.fit(embedding
) # pass node embeddings to be used to create edge embeddings
train_edges, train_labels = lpt.transform(
positive_graph=graphs['pos_training'],
negative_graph=graphs['neg_training'])
valid_edges, valid_labels = lpt.transform(
positive_graph=graphs['pos_validation'],
negative_graph=graphs['neg_validation'])
return (train_edges, train_labels), (valid_edges, valid_labels)
def _fit_transform(self,
graph: Graph,
return_dataframe: bool = True,
verbose: bool = True) -> EmbeddingResult:
"""Return node embedding."""
edges, weights = graph.get_log_normalized_cooccurrence_coo_matrix(
**self._walk_parameters)
coo = coo_matrix(
(weights, (edges[:, 0], edges[:, 1])),
shape=(graph.get_number_of_nodes(), graph.get_number_of_nodes()),
dtype=np.float32)
model = TruncatedSVD(n_components=self._embedding_size,
random_state=self._random_state)
model.fit(coo)
embedding = model.transform(coo)
if return_dataframe:
node_names = graph.get_node_names()
embedding = pd.DataFrame(embedding, index=node_names)
return EmbeddingResult(embedding_method_name=self.model_name(),
node_embeddings=embedding)
def _fit_transform(self,
graph: Graph,
return_dataframe: bool = True,
verbose: bool = True) -> EmbeddingResult:
"""Return node embedding."""
node_embedding, edge_type_embedding = self._model.fit_transform(
graph,
epochs=self._epochs,
learning_rate=self._learning_rate,
learning_rate_decay=self._learning_rate_decay,
verbose=verbose,
)
if return_dataframe:
node_embedding = pd.DataFrame(node_embedding,
index=graph.get_node_names())
edge_type_embedding = pd.DataFrame(
edge_type_embedding, index=graph.get_unique_edge_type_names())
return EmbeddingResult(
embedding_method_name=self.model_name(),
node_embeddings=node_embedding,
edge_type_embeddings=edge_type_embedding,
)
def _fit_transform(self,
graph: Graph,
return_dataframe: bool = True,
verbose: bool = True) -> EmbeddingResult:
"""Return node embedding."""
node_embedding = self._model.fit_transform(
graph,
verbose=verbose,
).T
if return_dataframe:
node_embedding = pd.DataFrame(node_embedding,
index=graph.get_node_names())
return EmbeddingResult(embedding_method_name=self.model_name(),
node_embeddings=node_embedding)
def _build_edge_prediction_based_model(
self, graph: Graph, sources: tf.Tensor,
destinations: tf.Tensor) -> Union[List[tf.Tensor], tf.Tensor]:
"""Return the model implementation.
Parameters
-------------------
sources: tf.Tensor
The source nodes to be used in the model.
destinations: tf.Tensor
The destinations nodes to be used in the model.
"""
node_embedding = Embedding(input_dim=graph.get_number_of_nodes(),
output_dim=self._embedding_size,
input_length=1,
name="node_embeddings")
context_embedding = Embedding(input_dim=graph.get_number_of_nodes(),
output_dim=self._embedding_size,
input_length=1,
name="context_embeddings")
return Activation(self._activation)(Dot(axes=-1)([
Flatten()(node_embedding(sources)),
Flatten()(context_embedding(destinations))
]))
def _extract_embeddings(self, graph: Graph, model: Model,
return_dataframe: bool) -> EmbeddingResult:
"""Returns embedding from the model.
Parameters
------------------
graph: Graph
The graph that was embedded.
model: Model
The Keras model used to embed the graph.
return_dataframe: bool
Whether to return a dataframe of a numpy array.
"""
if return_dataframe:
result = {
layer_name: pd.DataFrame(self.get_layer_weights(
layer_name, model, drop_first_row=drop_first_row),
index=names)
for layer_name, names, drop_first_row in (
("node_embeddings", graph.get_node_names(),
False), ("edge_type_embeddings",
graph.get_unique_edge_type_names(),
graph.has_unknown_edge_types()))
}
else:
result = {
layer_name:
self.get_layer_weights(layer_name,
model,
drop_first_row=drop_first_row)
for layer_name, drop_first_row in (
("node_embeddings", False),
("edge_type_embeddings", graph.has_unknown_edge_types()))
}
return EmbeddingResult(embedding_method_name=self.model_name(),
**result)
def _fit_transform(self,
graph: Graph,
return_dataframe: bool = True,
verbose: bool = True) -> EmbeddingResult:
"""Return node embedding."""
node_embeddings = self._model.fit_transform(graph)
if not isinstance(node_embeddings, list):
node_embeddings = [node_embeddings]
if return_dataframe:
node_names = graph.get_node_names()
node_embeddings = [
pd.DataFrame(node_embedding, index=node_names)
for node_embedding in node_embeddings
]
return EmbeddingResult(embedding_method_name=self.model_name(),
node_embeddings=node_embeddings)
def _fit_transform(
self,
graph: Graph,
return_dataframe: bool = True,
verbose: bool = True
) -> Union[np.ndarray, pd.DataFrame, Dict[str, np.ndarray], Dict[
str, pd.DataFrame]]:
"""Return node embedding"""
torch_device = torch.device(self._device)
triples_factory = CoreTriplesFactory(
torch.IntTensor(graph.get_directed_edge_triples_ids().astype(
np.int32)),
num_entities=graph.get_number_of_nodes(),
num_relations=graph.get_number_of_edge_types(),
entity_ids=graph.get_node_ids(),
relation_ids=graph.get_unique_edge_type_ids(),
create_inverse_triples=False,
)
batch_size = min(self._batch_size,
graph.get_number_of_directed_edges())
model = self._build_model(triples_factory)
if not issubclass(model.__class__, Model):
raise NotImplementedError(
"The model created with the `_build_model` in the child "
f"class {self.__class__.__name__} for the model {self.model_name()} "
f"in the library {self.library_name()} did not return a "
f"PyKeen model but an object of type {type(model)}.")
# Move the model to gpu if we need to
model.to(torch_device)
training_loop = SLCWATrainingLoop(
model=model,
triples_factory=triples_factory,
)
training_loop.train(triples_factory=triples_factory,
num_epochs=self._epochs,
batch_size=batch_size,
use_tqdm=True,
use_tqdm_batch=True,
tqdm_kwargs=dict(disable=not verbose))
# Extract and return the embedding
return self._extract_embeddings(graph,
model,
return_dataframe=return_dataframe)
def convert_graph_to_kernel(self,
graph: Graph) -> Optional[tf.SparseTensor]:
"""Returns provided graph converted to a sparse Tensor.
Implementation details
---------------------------
Do note that when the model does not have convolutional layers
the model will return None, as to avoid allocating like object for
apparently no reason.
"""
if not self.has_convolutional_layers():
return None
return graph_to_sparse_tensor(
graph,
use_weights=graph.has_edge_weights()
and not self._use_simmetric_normalized_laplacian,
use_simmetric_normalized_laplacian=self.
_use_simmetric_normalized_laplacian,
handling_multi_graph=self._handling_multi_graph)
def _fit_transform(self,
graph: Graph,
return_dataframe: bool = True,
verbose: bool = True) -> EmbeddingResult:
"""Return node embedding.
Parameters
---------------
graph: Graph
The graph to embed.
return_dataframe: bool = True
Whether to return a DataFrame.
verbose: bool = True
Whether to show a loading bar.
"""
model: Type[Estimator] = self._build_model()
if not issubclass(model.__class__, Estimator):
raise NotImplementedError(
"The model created with the `_build_model` in the child "
f"class {self.__class__.__name__} for the model {self.model_name()} "
f"in the library {self.library_name()} did not return a "
f"Estimator but an object of type {type(model)}. "
"It is not clear what to do with this object.")
model.fit(convert_ensmallen_graph_to_networkx_graph(graph))
node_embeddings: np.ndarray = model.get_embedding()
if not issubclass(node_embeddings.__class__, np.ndarray):
raise NotImplementedError(
"The model created with the `get_embedding` in the child "
f"class {self.__class__.__name__} for the model {self.model_name()} "
f"in the library {self.library_name()} did not return a "
f"Numpy Array but an object of type {type(model)}. "
"It is not clear what to do with this object.")
if return_dataframe:
node_embeddings: pd.DataFrame = pd.DataFrame(
node_embeddings, index=graph.get_node_names())
return EmbeddingResult(embedding_method_name=self.model_name(),
node_embeddings=node_embeddings)
