Graph Transforms

k3_node.transforms provides 62+ deterministic and stochastic transformations for graphs, 3D point clouds, meshes, and node features.

Transforms can be chained using Compose and passed to datasets as transform (applied dynamically on each get) or pre_transform (applied once during offline processing).

from k3_node.transforms import Compose, ToUndirected, AddSelfLoops, NormalizeFeatures

transform = Compose([
    ToUndirected(),
    AddSelfLoops(),
    NormalizeFeatures(),
])

data = transform(data)

General Transforms

Transforms operating on node/edge feature matrices, masks, and splits:

Transform	Description
`Compose`	Sequentially chains multiple transforms together.
`ComposeFilters`	Filters graphs based on Boolean predicate functions.
`NormalizeFeatures`	Row-normalizes node features (\(L_1\) or \(L_2\) norm).
`Constant`	Adds a constant scalar or vector value to node features.
`RandomNodeSplit`	Generates train/validation/test node masks randomly or with fixed counts.
`RandomLinkSplit`	Splits edge indices into train/val/test edges with negative sampling.
`NodePropertySplit`	Partitions nodes based on continuous or categorical properties.
`IndexToMask`	Converts an array of indices into a Boolean mask.
`MaskToIndex`	Converts a Boolean mask into an array of active indices.
`SVDFeatureReduction`	Dimensionality reduction on node feature matrices via SVD.
`RemoveTrainingClasses`	Removes specific classes from the training set for open-world settings.
`Pad`	Pads node feature matrices and adjacency representations to a fixed budget.
`ToSparseTensor`	Converts `edge_index` to sparse adjacency matrix format.
`ToDevice`	Transfers tensor attributes to a specific compute device.

`k3_node.transforms.Compose`

Bases: BaseTransform

Composes several transforms together.

Parameters:

Name	Type	Description	Default
`transforms`	`List[Callable]`	List of transforms to compose.	required

`k3_node.transforms.NormalizeFeatures`

Bases: BaseTransform

Row-normalizes node features to sum to 1 (L1-norm).

Graph Structural & Spectral Transforms

Transforms modifying graph topology, adding self-loops, computing positional encodings, and rewiring edges:

Transform	Description
`ToUndirected`	Converts a directed graph into an undirected graph by adding reciprocal edges.
`AddSelfLoops`	Adds self-loops \((i, i)\) to all nodes.
`AddRemainingSelfLoops`	Adds self-loops only to nodes that do not already have one.
`RemoveSelfLoops`	Removes all self-loops from the graph.
`RemoveIsolatedNodes`	Prunes isolated nodes without any incident edges.
`RemoveDuplicatedEdges`	Deduplicates multi-edges in the graph.
`TwoHop`	Adds edges connecting nodes that are 2 hops apart.
`LineGraph`	Constructs the line graph where original edges become nodes.
`LargestConnectedComponents`	Restricts the graph to its largest connected component(s).
`VirtualNode`	Adds an auxiliary virtual super-node connected to all nodes.
`GCNNorm`	Computes standard GCN symmetric normalized adjacency matrix \(\mathbf{\tilde{D}}^{-1/2}\mathbf{\tilde{A}}\mathbf{\tilde{D}}^{-1/2}\).
`GDC`	Generalized Graph Diffusion (PageRank or Heat kernel diffusion) for graph denoising and rewiring.
`SIGN`	Precomputes multi-hop diffused feature aggregations for Scalable Inception Graph Neural Networks.
`AddLaplacianEigenvectorPE`	Computes Laplacian eigenvector positional encodings (\(k\) smallest non-trivial eigenvectors).
`AddRandomWalkPE`	Computes Random Walk structural positional encodings (landing probabilities \(RW_{i,i}^k\)).
`AddGPSE`	Computes 20-dimensional Graph Positional and Structural Encodings.
`FeaturePropagation`	Reconstructs missing node features via Dirichlet energy diffusion.
`HalfHop`	Slow-fast graph rewiring for over-smoothing and over-squashing mitigation.
`OneHotDegree`	Computes one-hot encodings of node degrees.
`TargetIndegree`	Appends target node in-degrees as edge features.
`LocalDegreeProfile`	Extracts statistical degree distribution summaries for local neighborhoods.
`KNNGraph`	Constructs a \(k\)-nearest-neighbor graph from spatial coordinates.
`RadiusGraph`	Constructs an \(\epsilon\)-ball radius graph from spatial coordinates.
`ToDense`	Converts sparse graph representation into dense adjacency and feature matrices.

`k3_node.transforms.ToUndirected`

Bases: BaseTransform

Converts a homogeneous or heterogeneous graph to an undirected graph.

`k3_node.transforms.AddSelfLoops`

Bases: BaseTransform

Adds self-loops to the graph.

`k3_node.transforms.AddLaplacianEigenvectorPE`

Bases: BaseTransform

Adds Laplacian eigenvector positional encoding.

`k3_node.transforms.AddRandomWalkPE`

Bases: BaseTransform

Adds random walk positional encoding.

Spatial & Geometric Transforms

Transforms for 3D point clouds, molecular coordinates, and triangle meshes:

Transform	Description
`Cartesian`	Saves relative Cartesian coordinates \(\mathbf{p}_j - \mathbf{p}_i\) as edge attributes.
`LocalCartesian`	Relative Cartesian coordinates projected onto local coordinate frames.
`Distance`	Computes Euclidean distances \(\\|\mathbf{p}_j - \mathbf{p}_i\\|_2\) as edge attributes.
`Polar`	Converts 2D/3D relative offsets to polar/cylindrical coordinates.
`Spherical`	Converts 3D relative offsets to spherical coordinates \((r, \theta, \phi)\).
`PointPairFeatures`	Computes Point Pair Features (PPF) using surface normal vectors.
`Center`	Centers coordinates around origin \((\sum \mathbf{p} = \mathbf{0})\).
`NormalizeScale`	Rescales coordinates into the unit sphere \([-1, 1]\).
`NormalizeRotation`	Aligns principal axes of point clouds via eigenvectors of inertia tensor.
`RandomRotate`	Randomly rotates 3D point cloud / mesh coordinates.
`RandomScale`	Randomly scales 3D coordinates by a factor sampled uniformly from \([\min, \max]\).
`RandomTranslate`	Randomly offsets coordinates with bounded translation noise.
`RandomJitter`	Adds Gaussian or uniform jitter noise to coordinates.
`RandomFlip`	Randomly mirrors coordinates along chosen axes.
`SamplePoints`	Uniformly samples a fixed number of points from triangle meshes.
`FixedPoints`	Subsamples or resamples point clouds to a fixed number of points.
`GenerateMeshNormals`	Computes per-face and per-vertex surface normal vectors from triangle faces.
`FaceToEdge`	Converts mesh triangular faces `face` into undirected edge indices `edge_index`.
`Delaunay`	Computes Delaunay triangulation for 2D/3D point clouds.
`GridSampling`	Voxel grid spatial downsampling of point clouds.
`ToSLIC`	Generates superpixels from images using Simple Linear Iterative Clustering.