dgf.data

dgf.data.EdgeSchema

dgf.data.EdgeSetPadding

dgf.data.FeatureFormat

Bases: Enum

How a value is represented / stored.

dgf.data.FeatureSchema

Schema for a single feature.

For GNN training, users are expected to provide the format (and possibly semantic if using auto-normalization) fields. Other fields are automatically inferred when computing feature statistics + (auto) normalization.

Attributes:

Name	Type	Description
`format`	`FeatureFormat`	The physical format of the feature values (e.g., INTEGER, FLOAT).
`semantic`	`FeatureSemantic`	The semantic meaning of the feature (e.g., NUMERICAL, CATEGORICAL).
`shape`	`Shape`	The shape of the feature tensor. [] or [1] for scalar features. `None` if the shape is unknown--but assumed to be () in some locations. Shape values None (e.g., [None]) indicates an unknown shape. Unknown shapes are not handelled by in-memory-graphs and TF-GNN Graph Samples.
`num_categorical_values`	`Optional[int]`	The number of possible categories for CATEGORICAL features. `None` for other semantic types or if the numeber of possible categories is unknown.
`is_utf8_string`	`Optional[bool]`	Whether the feature is a UTF-8 string. This is only relevant when feature_format is BYTES, to distinguish between Spanner STRING (True) and Spanner BYTES (False).
`is_timeseries`	`bool`	Whether the feature represents a temporal series / sequence.
`is_creation_time`	`bool`	Whether the feature represents a creation time. For non-timeseries features (`is_timeseries=False`), it indicates the creation time of the node or edge entity itself. For timeseries features (`is_timeseries=True`), it marks the sequence as the reference timestamp sequence for all features in the same `group`. This information is used used in temporal neighbor sampling and for masking future information in timeseries.
`group`	`Optional[str]`	Name of the sequence group this feature belongs to. All sequence features sharing the same `group` share the same sequence length along `shape[0]` and share a single optional group mask and optional group timestamp sequence.

`is_static_shape`

Returns true if the feature has a fully static shape.

`static_size`

Product of all the static dimeneions.

dgf.data.FeatureSemantic

Bases: Enum

How a value should be interpreted.

Possible values

NUMERICAL: The feature represents a numerical quantity e.g. count, amount. CATEGORICAL: The feature represents a category from a finite set of values without ordering. EMBEDDING: The feature is a dense and normalized vector embedding that can be directly consumed by a neural network. TIMESTAMP: The feature represents a timestamp. TIMEDELTA: The feature represents a time delta or time duration. TIMESERIES: The feature represents a time series of values. PRIMARY_ID: The feature represents a primary ID. MASK: The feature represents a boolean sequence padding mask. It must belong to a timeseries_group, and its shape has to be broadcastable to the features in that group.

dgf.data.FeatureSetStatistics

Statistics for a set of features.

dgf.data.FeatureStatistics

Statistics for a feature.

Attributes:

Name	Type	Description
`count`	`int`	The number of values for the feature. For multi-dim features, each observation counts as 1.
`minimum`	`float`	Minimum value. For INTERGER and FLOAT only.
`maximum`	`float`	Maximum value. For INTERGER and FLOAT only.
`dictionary`	`Dictionary`	Dictionary of key->(index, count). For BYTES only. Notes: Non-utf8 characters are encoded using the 'surrogateescape' method.
`quantiles`	`List[float]`	Quantiles over the dataset distribution.

dgf.data.GraphFeatureStatistics

Statistics about the features in a graph.

dgf.data.GraphSchema

dgf.data.GraphSchemaFilter

Filters a GraphSchema to sub-select node sets, edge sets, and features.

A GraphSchemaFilter can be used with IO methods (when reading a graph from disk; with the schema_filter argument when available), or to operate directly on a schema (with dgf.transform.filter_schema).

Attributes:

Name	Type	Description
`nodeset_fn`	`Callable[[str, NodeSchema], bool]`	A callable `(name, schema) -> bool` to filter node sets. If it returns True, keep the nodeset. Otherwise, remove it.
`edgeset_fn`	`Callable[[str, EdgeSchema], bool]`	A callable `(name, schema) -> bool` to filter edge sets. If it returns True, keep the edgeset. Otherwise, remove it.
`feature_fn`	`Callable[[str, FeatureSchema], bool]`	A callable `(name, schema) -> bool` to filter node OR edge features. If it returns True, keep the feature. Otherwise, remove it.
`node_feature_fn`	`Callable[[str, FeatureSchema], bool]`	A callable `(name, schema) -> bool` to filter node features. If it returns True, keep the feature. Otherwise, remove it.
`edge_feature_fn`	`Callable[[str, FeatureSchema], bool]`	A callable `(name, schema) -> bool` to filter edge features. If it returns True, keep the feature. Otherwise, remove it.

dgf.data.GraphSchemaV2

dgf.data.InMemoryEdgeSet

An Edge Set.

Attributes:

Name	Type	Description
`adjacency`	`ndarray`	Array of shape [2, num_edges]. Dtype: int-like. If the graph is batched, the batch dimension is added first: [batch, num_edges].
`features`	`Features`	Dictionary of edge set features. Feature values are ArrayOrAnys of shape [num_edges, ].

dgf.data.InMemoryGraph

An in-memory generic graph.

Attributes:

Name	Type	Description
`node_sets`	`Dict[str, InMemoryNodeSet]`	Dictionary of node set name and node set features.
`edge_sets`	`Dict[str, InMemoryEdgeSet]`	Dictionary of edge set name and edge set features.

dgf.data.InMemoryNodeSet

A Node Set.

Attributes:

Name	Type	Description
`num_nodes`	`Optional[int]`	Number of nodes in the nodeset. Useful to determine the number of nodes in a nodeset without features.
`features`	`Features`	Dictionary of feature names to feature values. Feature values are ArrayOrAnys of shape [num_nodes, ].

dgf.data.JaxInMemoryEdgeSet

An Edge Set.

Attributes:

Name	Type	Description
`adjacency`	`Array`	Array of shape [2, num_edges]. Dtype: int-like. If the graph is batched, the batch dimension is added first: [batch, num_edges].
`features`	`Features`	Dictionary of edge set features. Feature values are ArrayOrAnys of shape [num_edges, ...]. If the graph is batched, the batch dimension is added first: [batch, num_edges, ...].

dgf.data.JaxInMemoryGraph

An in-memory generic graph.

Attributes:

Name	Type	Description
`node_sets`	`Dict[str, JaxInMemoryNodeSet]`	Dictionary of node set name and node set features.
`edge_sets`	`Dict[str, JaxInMemoryEdgeSet]`	Dictionary of edge set name and edge set features.

dgf.data.JaxInMemoryNodeSet

A Node Set.

Attributes:

Name	Type	Description
`features`	`Features`	Dictionary of feature names to feature values. Feature values are ArrayOrAnys of shape [num_nodes, ...]. If the graph is batched, the batch dimension is added first: [batch, num_nodes, ...].
`num_nodes`	`Optional[int]`	Number of nodes in the nodeset. Useful to determine the number of nodes in a nodeset without features.

dgf.data.NodeSchema

dgf.data.NodeSetPadding

dgf.data.Padding

Information to pad a graph.

dgf.data.TFInMemoryEdgeSet

Bases: ExtensionType

An Edge Set.

Attributes:

Name	Type	Description
`adjacency`	`Array`	Array of shape [2, num_edges]. Dtype: int-like. If the graph is batched, the batch dimension is added first: [batch, num_edges].
`features`	`Features`	Dictionary of edge set features. Feature values are ArrayOrAnys of shape [num_edges, ].

dgf.data.TFInMemoryGraph

Bases: ExtensionType

An in-memory generic graph.

Attributes:

Name	Type	Description
`node_sets`	`Mapping[str, TFInMemoryNodeSet]`	Dictionary of node set name and node set features.
`edge_sets`	`Mapping[str, TFInMemoryEdgeSet]`	Dictionary of edge set name and edge set features.

dgf.data.TFInMemoryNodeSet

Bases: ExtensionType

A Node Set.

Attributes:

Name	Type	Description
`num_nodes`	`Union[int, Array]`	Number of nodes in the nodeset. Useful to determine the number of nodes in a nodeset without features.
`features`	`Features`	Dictionary of feature names to feature values. Feature values are ArrayOrAnys of shape [num_nodes, ].