k3-node Porting Checklist

Tracks parity between torch_geometric.nn and k3_node.layers / k3_node.models. Generated by diffing torch_geometric.nn.<submodule>.__all__ against the equivalent k3-node package. Update this file whenever new API is ported.

Legend: ✅ ported · 🟡 partially ported · ⬜ not ported

Summary

Submodule	k3-node location	Status
torch_geometric.nn.conv	k3_node.layers.conv	✅ 65/65
torch_geometric.nn.pool	k3_node.layers.pool	✅ 31/31
torch_geometric.nn.unpool	k3_node.layers.unpool	✅ 1/1
torch_geometric.nn.kge	k3_node.layers.kge	✅ 5/5
torch_geometric.nn.functional	k3_node.layers.functional	✅ 2/2
torch_geometric.nn.norm	k3_node.layers.norm	✅ 11/11
torch_geometric.nn.aggr	k3_node.layers.aggr	✅ 26/26
torch_geometric.nn.dense	k3_node.layers.dense	✅ 11/11
torch_geometric.nn.models	k3_node.models	✅ 46/46
torch_geometric.data	k3_node.data	✅ 19/19
torch_geometric.loader	k3_node.loader	✅ 26/26
torch_geometric.transforms	k3_node.transforms	✅ 62/62

conv, pool, unpool, kge, functional, norm, aggr, dense, models, data, loader, and transforms are at full parity with PyG's public API (see each package's __all__).

torch_geometric.nn.models (parity status)

PyG class	Status	k3-node location	Notes
MLP	✅	k3_node.models.MLP
ARLinkPredictor	✅	k3_node.models.ARLinkPredictor	Attract-Repel link predictor
InnerProductDecoder	✅	k3_node.models.InnerProductDecoder
GAE	✅	k3_node.models.GAE
VGAE	✅	k3_node.models.VGAE
ARGA	✅	k3_node.models.ARGA
ARGVA	✅	k3_node.models.ARGVA
DeepGraphInfomax	✅	k3_node.models.DeepGraphInfomax
DeepGCNLayer	✅	k3_node.models.DeepGCNLayer	ckpt_grad accepted but a no-op (no uniform gradient-checkpoint API across Keras backends)
AttentiveFP	✅	k3_node.models.AttentiveFP	Molecular graph attention model
GCN	✅	k3_node.models.GCN	BasicGNN with GCNConv layers
GraphSAGE	✅	k3_node.models.GraphSAGE	BasicGNN with SAGEConv layers
GIN	✅	k3_node.models.GIN	BasicGNN with GINConv layers + MLP
GAT	✅	k3_node.models.GAT	BasicGNN with GATConv/GATv2Conv layers
PNA	✅	k3_node.models.PNA	BasicGNN with PNAConv layers
EdgeCNN	✅	k3_node.models.EdgeCNN	BasicGNN with EdgeConv layers + MLP
JumpingKnowledge	✅	k3_node.models.JumpingKnowledge	Supports cat, max, and lstm modes
HeteroJumpingKnowledge	✅	k3_node.models.HeteroJumpingKnowledge	Heterogeneous graph jumping knowledge
MetaLayer	✅	k3_node.models.MetaLayer	Edge, node, and global network modules
Node2Vec	✅	k3_node.models.Node2Vec	Random-walk + skip-gram embedding without pyg-lib C++ dependency
SignedGCN	✅	k3_node.models.SignedGCN	Signed graph convolutional network with balanced losses
RENet	✅	k3_node.models.RENet	Temporal knowledge graph reasoning with GRU and mean aggregation
GraphUNet	✅	k3_node.models.GraphUNet	U-Net-style pooling/unpooling GNN
SchNet	✅	k3_node.models.SchNet	Molecular continuous-filter convs with GaussianSmearing & ShiftedSoftplus
DimeNet / DimeNetPlusPlus	✅	k3_node.models.DimeNet / DimeNetPlusPlus	Directional message passing with BesselBasis, SphericalBasis, and triplets
GPSE / GPSENodeEncoder	✅	k3_node.models.GPSE / GPSENodeEncoder	Graph Positional & Structural Encoder with multi-head predictions & node encoder
to_captum_model / to_captum_input / captum_output_to_dicts	✅	k3_node.models.captum	PyTorch-exclusive interpretability stubs with documentation
MetaPath2Vec	✅	k3_node.models.MetaPath2Vec	Heterogeneous network embedding with metapath random walks
TGNMemory	✅	k3_node.models.TGNMemory	Temporal Graph Network memory model with TimeEncoder & aggregators
LabelPropagation	✅	k3_node.models.LabelPropagation	Label propagation via symmetric normalized Laplacian
CorrectAndSmooth	✅	k3_node.models.CorrectAndSmooth	Residual error correction + label smoothing
RECT_L	✅	k3_node.models.RECT_L	Supervised class-reconstruction model
LINKX	✅	k3_node.models.LINKX	Non-homophilous graph learner with SparseLinear
LightGCN	✅	k3_node.models.LightGCN	Light graph convolution network for recommendation
MaskLabel	✅	k3_node.models.MaskLabel	Label embedding and masking layer (add and concat)
GroupAddRev	✅	k3_node.models.GroupAddRev	Reversible GNN block with inverse
GNNFF	✅	k3_node.models.GNNFF	Graph Neural Network Force Field with NodeBlock, EdgeBlock, and GaussianFilter
PMLP	✅	k3_node.models.PMLP	Propagational MLP with SimpleConv propagation at inference
NeuralFingerprint	✅	k3_node.models.NeuralFingerprint	Molecular neural fingerprint with MFConv and add pool
ViSNet	✅	k3_node.models.ViSNet	Equivariant vector-scalar interactive GNN with spherical harmonics & CosineCutoff
LPFormer	✅	k3_node.models.LPFormer	Adaptive graph transformer for link prediction with PPR positional encodings
SGFormer	✅	k3_node.models.SGFormer	All-pair linear attention + GNN graph transformer
Polynormer	✅	k3_node.models.Polynormer	Full polynomial-expressive graph transformer (local + global)

torch_geometric.data (parity status)

PyG class / function	Status	k3-node location	Notes
Data	✅	k3_node.data.Data	Homogeneous graph container with multi-backend tensor support
HeteroData	✅	k3_node.data.HeteroData	Heterogeneous graph container with multi-backend tensor support
Batch / HeteroBatch	✅	k3_node.data.Batch	Disjoint graph batching & separation for Data & HeteroData
TemporalData	✅	k3_node.data.TemporalData	Continuous-time temporal event stream container
HypergraphData	✅	k3_node.data.HypergraphData	Hypergraph container with hyperedge indexing
Dataset	✅	k3_node.data.Dataset	Base graph dataset abstraction
InMemoryDataset	✅	k3_node.data.InMemoryDataset	In-memory graph dataset with serialization
OnDiskDataset	✅	k3_node.data.OnDiskDataset	Out-of-core graph dataset with database backend
FeatureStore	✅	k3_node.data.FeatureStore	Multi-backend feature store interface
GraphStore	✅	k3_node.data.GraphStore	Graph edge store interface
TensorAttr	✅	k3_node.data.TensorAttr	Feature store tensor metadata attribute
EdgeAttr / EdgeLayout	✅	k3_node.data.EdgeAttr	Graph store edge layout and metadata
Database	✅	k3_node.data.Database	Abstract key-value database interface
SQLiteDatabase	✅	k3_node.data.SQLiteDatabase	SQLite3-backed persistent graph store
RocksDatabase	✅	k3_node.data.RocksDatabase	RocksDB-backed database stub
makedirs	✅	k3_node.data.makedirs	Recursive directory creation utility
download_url	✅	k3_node.data.download_url	HTTP file downloader with progress
download_google_url	✅	k3_node.data.download_google_url	Google Drive download helper
extract_tar / extract_zip / extract_bz2 / extract_gz	✅	k3_node.data.extract_*	Archive extraction helpers

torch_geometric.loader (parity status)

PyG class	Status	k3-node location	Notes
DataLoader	✅	k3_node.loader.DataLoader	Multi-backend graph batch loader with follow_batch & exclude_keys
NodeLoader	✅	k3_node.loader.NodeLoader	Mini-batch sampling from node information
LinkLoader	✅	k3_node.loader.LinkLoader	Mini-batch sampling from link/edge information
NeighborLoader	✅	k3_node.loader.NeighborLoader	Multi-hop neighbor sampling without external C++ binary dependency
LinkNeighborLoader	✅	k3_node.loader.LinkNeighborLoader	Link-centric neighbor sampling with positive & negative link supervision
HGTLoader	✅	k3_node.loader.HGTLoader	Heterogeneous Graph Transformer balanced neighbor sampling
ClusterData	✅	k3_node.loader.ClusterData	Graph partitioner with METIS and pure Python/BFS fallback
ClusterLoader	✅	k3_node.loader.ClusterLoader	Merges partitioned subgraphs into mini-batches
GraphSAINTSampler	✅	k3_node.loader.GraphSAINTSampler	GraphSAINT random subgraph sampler base class
GraphSAINTNodeSampler	✅	k3_node.loader.GraphSAINTNodeSampler	Node-based GraphSAINT sampler
GraphSAINTEdgeSampler	✅	k3_node.loader.GraphSAINTEdgeSampler	Edge-probability GraphSAINT sampler
GraphSAINTRandomWalkSampler	✅	k3_node.loader.GraphSAINTRandomWalkSampler	Random-walk-based GraphSAINT sampler
ShaDowKHopSampler	✅	k3_node.loader.ShaDowKHopSampler	Decoupled shallow ego-network subgraph extractor
RandomNodeLoader	✅	k3_node.loader.RandomNodeLoader	Random node partition loader for large graphs
ZipLoader	✅	k3_node.loader.ZipLoader	Combines multiple loaders into synchronized tuples
DataListLoader	✅	k3_node.loader.DataListLoader	Yields lists of Data objects without batch merging
DenseDataLoader	✅	k3_node.loader.DenseDataLoader	Stacks graphs with dense adjacency matrices
TemporalDataLoader	✅	k3_node.loader.TemporalDataLoader	Successive temporal event window mini-batch loader
NeighborSampler	✅	k3_node.loader.NeighborSampler	Legacy layer-by-layer bipartite neighbor sampler
ImbalancedSampler	✅	k3_node.loader.ImbalancedSampler	Class-frequency weighted random sampler
DynamicBatchSampler	✅	k3_node.loader.DynamicBatchSampler	Dynamic node/edge budget mini-batch sampler
PrefetchLoader	✅	k3_node.loader.PrefetchLoader	Asynchronous host-to-device memory prefetcher
CachedLoader	✅	k3_node.loader.CachedLoader	In-memory mini-batch cache across epochs
AffinityMixin	✅	k3_node.loader.AffinityMixin	CPU worker core affinitization context manager
MultithreadingMixin	✅	k3_node.loader.MultithreadingMixin	Worker subprocess thread count configuration
LogMemoryMixin	✅	k3_node.loader.LogMemoryMixin	Worker RSS memory consumption logger

torch_geometric.transforms (parity status)

Group	PyG classes	Status	k3-node location	Notes
General (15)	BaseTransform, Compose, ComposeFilters, ToDevice, ToSparseTensor, Constant, NormalizeFeatures, SVDFeatureReduction, RemoveTrainingClasses, RandomNodeSplit, RandomLinkSplit, NodePropertySplit, IndexToMask, MaskToIndex, Pad	✅ 15/15	k3_node.transforms.general	Multi-backend support (torch.Tensor, numpy arrays, Keras tensors)
Graph (26)	ToUndirected, OneHotDegree, TargetIndegree, LocalDegreeProfile, AddSelfLoops, AddRemainingSelfLoops, RemoveSelfLoops, RemoveIsolatedNodes, RemoveDuplicatedEdges, KNNGraph, RadiusGraph, ToDense, TwoHop, LineGraph, LaplacianLambdaMax, GDC, SIGN, GCNNorm, AddMetaPaths, AddRandomMetaPaths, RootedEgoNets, RootedRWSubgraph, LargestConnectedComponents, VirtualNode, AddLaplacianEigenvectorPE, AddRandomWalkPE, AddGPSE, FeaturePropagation, HalfHop	✅ 26/26	k3_node.transforms.graph	Self-contained graph transformations with sparse matrix computations
Vision / Spatial (21)	Distance, Cartesian, LocalCartesian, Polar, Spherical, PointPairFeatures, Center, NormalizeRotation, NormalizeScale, RandomJitter, RandomFlip, LinearTransformation, RandomScale, RandomRotate, RandomShear, FaceToEdge, SamplePoints, FixedPoints, GenerateMeshNormals, Delaunay, ToSLIC, GridSampling, RandomTranslate	✅ 21/21	k3_node.transforms.spatial	3D mesh and point cloud geometric processing with Delaunay & grid clustering

Notes on non-goals / caveats

• k3_node.layers.kge.KGTripletLoader replaces PyG's torch.utils.data.DataLoader-based loader with a framework-agnostic batching iterator (numpy-backed), since a torch-specific DataLoader doesn't fit a multi-backend (TF/PyTorch/JAX) library.
• k3_node.layers.unpool.knn_interpolate reuses k3_node.layers.pool.knn (a self-contained numpy brute-force k-NN) rather than PyG's pyg-lib/ torch-cluster-backed extension.
• Methods that are inherently framework-agnostic-unfriendly (e.g. PyTorch's torch.utils.checkpoint, TorchScript export, Captum integration) are ported as no-ops or omitted, with notes above.