Convolution Layers

The k3_node.layers.conv module provides over 65 spatial, spectral, relational, and point-cloud graph convolution layers.

Message Passing Base Class

`k3_node.layers.conv.MessagePassing`

Bases: Layer

Base class for creating Message Passing Neural Networks (MPNNs).

Parameters:

Name	Type	Description	Default
`aggr`	`Union[str, List[str], Aggregation, None]`	The aggregation scheme to use, such as `"add"`, `"sum"`, `"mean"`, `"min"`, `"max"`, `"mul"`, or an instance of :class:`~k3_node.layers.aggr.Aggregation`. (default: `"add"`)	`'add'`
`flow`	`str`	The direction of message passing (`"source_to_target"` or `"target_to_source"`). (default: `"source_to_target"`)	`'source_to_target'`
`node_dim`	`int`	The axis along which to index node features. (default: `-2`)	`-2`
`decomposed_layers`	`int`	Number of decomposed layers for memory-efficient aggregation. (default: `1`)	`1`

`aggregate(inputs=None, index=None, ptr=None, dim_size=None, **kwargs)`

Aggregates messages from neighbors as given by :obj:index.

`call(inputs, edge_index=None, **kwargs)`

Default call handler supporting both (x, edge_index) and legacy (inputs,) tuples.

`edge_update(**kwargs)`

Computes or updates edge attributes.

`edge_updater(edge_index, size=None, **kwargs)`

Computes or updates edge-level representations.

`message(x=None, x_j=None, **kwargs)`

Constructs messages from node :math:j to node :math:i.

`propagate(*args, **kwargs)`

The initial call to start propagating messages.

`update(embeddings=None, **kwargs)`

Updates node embeddings.

Core Graph Convolutions

GCNConv

`k3_node.layers.conv.GCNConv`

Bases: MessagePassing

The graph convolutional operator from the "Semi-supervised Classification with Graph Convolutional Networks" <https://arxiv.org/abs/1609.02907>_ paper.

.. math:: \mathbf{X}^{\prime} = \mathbf{\hat{D}}^{-1/2} \mathbf{\hat{A}} \mathbf{\hat{D}}^{-1/2} \mathbf{X} \mathbf{\Theta}

Parameters:

Name	Type	Description	Default
`in_channels`	`int`	Size of each input sample.	required
`out_channels`	`int`	Size of each output sample.	required
`improved`	`bool`	If set to :obj:`True`, the layer computes :math:`\mathbf{\hat{A}} = \mathbf{A} + 2 \mathbf{I}`. (default: :obj:`False`)	`False`
`cached`	`bool`	If set to :obj:`True`, the layer will cache the computation of :math:`\mathbf{\hat{D}}^{-1/2} \mathbf{\hat{A}} \mathbf{\hat{D}}^{-1/2}`. (default: :obj:`False`)	`False`
`add_self_loops`	`bool`	If set to :obj:`False`, will not add self-loops to the input graph. (default: :obj:`True`)	`True`
`normalize`	`bool`	Whether to add self-loops and compute symmetric normalization coefficients on the fly. (default: :obj:`True`)	`True`
`bias`	`bool`	If set to :obj:`False`, the layer will not learn an additive bias. (default: :obj:`True`)	`True`

SAGEConv

`k3_node.layers.conv.SAGEConv`

Bases: MessagePassing

The GraphSAGE operator from the "Inductive Representation Learning on Large Graphs" <https://arxiv.org/abs/1706.02216>_ paper.

Parameters:

Name	Type	Description	Default
`in_channels`	`Union[int, Tuple[int, int], None]`	Size of each input sample, or a tuple for bipartite graphs.	required
`out_channels`	`Optional[int]`	Size of each output sample.	`None`
`aggr`	`Optional[Union[str, List[str], Aggregation]]`	The aggregation scheme to use (`"mean"`, `"max"`, `"lstm"`, etc.). (default: `"mean"`)	`'mean'`
`normalize`	`bool`	If set to :obj:`True`, output features will be :math:`\ell_2`-normalized. (default: :obj:`False`)	`False`
`root_weight`	`bool`	If set to :obj:`False`, the layer will not add the transformed root node features. (default: :obj:`True`)	`True`
`project`	`bool`	If set to :obj:`True`, the layer will apply a linear transformation followed by an activation to source node features. (default: :obj:`False`)	`False`
`bias`	`bool`	If set to :obj:`False`, the layer will not learn an additive bias. (default: :obj:`True`)	`True`

GATConv

`k3_node.layers.conv.GATConv`

Bases: MessagePassing

The graph attentional operator from the "Graph Attention Networks" <https://arxiv.org/abs/1710.10903>_ paper.

Parameters:

Name	Type	Description	Default
`in_channels`	`Union[int, Tuple[int, int]]`	Size of each input sample, or a tuple for bipartite graphs.	required
`out_channels`	`int`	Size of each output sample.	required
`heads`	`int`	Number of multi-head-attentions. (default: `1`)	`1`
`concat`	`bool`	If set to :obj:`False`, the multi-head-attentions are averaged instead of concatenated. (default: `True`)	`True`
`negative_slope`	`float`	LeakyReLU angle of the negative slope. (default: `0.2`)	`0.2`
`dropout`	`float`	Dropout probability of the normalized attention coefficients. (default: `0.0`)	`0.0`
`add_self_loops`	`bool`	If set to :obj:`False`, will not add self-loops to the input graph. (default: `True`)	`True`
`edge_dim`	`Optional[int]`	Edge feature dimensionality (in case there are any). (default: :obj:`None`)	`None`
`fill_value`	`Union[float, str]`	The way to generate edge features of self-loops (default: `"mean"`)	`'mean'`
`bias`	`bool`	If set to :obj:`False`, the layer will not learn an additive bias. (default: `True`)	`True`
`share_weights`	`bool`	If set to :obj:`True`, the same matrix will be applied to the source and target node features. (default: `False`)	`False`
`residual`	`bool`	If set to :obj:`True`, will compute residual connections. (default: `False`)	`False`

GATv2Conv

`k3_node.layers.conv.GATv2Conv`

Bases: MessagePassing

The GATv2 operator from the "How Attentive are Graph Attention Networks?" <https://arxiv.org/abs/2105.14491>_ paper, which fixes the static attention problem of standard :class:~k3_node.layers.conv.GATConv.

Parameters:

Name	Type	Description	Default
`in_channels`	`Union[int, Tuple[int, int]]`	Size of each input sample, or a tuple for bipartite graphs.	required
`out_channels`	`int`	Size of each output sample.	required
`heads`	`int`	Number of multi-head-attentions. (default: `1`)	`1`
`concat`	`bool`	If set to :obj:`False`, the multi-head-attentions are averaged instead of concatenated. (default: `True`)	`True`
`negative_slope`	`float`	LeakyReLU angle of the negative slope. (default: `0.2`)	`0.2`
`dropout`	`float`	Dropout probability of the normalized attention coefficients. (default: `0.0`)	`0.0`
`add_self_loops`	`bool`	If set to :obj:`False`, will not add self-loops to the input graph. (default: `True`)	`True`
`edge_dim`	`Optional[int]`	Edge feature dimensionality (in case there are any). (default: :obj:`None`)	`None`
`fill_value`	`Union[float, str]`	The way to generate edge features of self-loops (default: `"mean"`)	`'mean'`
`bias`	`bool`	If set to :obj:`False`, the layer will not learn an additive bias. (default: `True`)	`True`
`share_weights`	`bool`	If set to :obj:`True`, the same matrix will be applied to the source and target node features. (default: `False`)	`False`
`residual`	`bool`	If set to :obj:`True`, will compute residual connections. (default: `False`)	`False`

TransformerConv

`k3_node.layers.conv.TransformerConv`

Bases: MessagePassing

The graph transformer operator from the "Masked Label Prediction: Unified Meta-Learning on Graph Neural Networks" <https://arxiv.org/abs/2009.03509>_ paper.

Parameters:

Name	Type	Description	Default
`in_channels`	`Union[int, Tuple[int, int]]`	Size of each input sample, or a tuple for bipartite graphs.	required
`out_channels`	`int`	Size of each output sample.	required
`heads`	`int`	Number of multi-head-attentions. (default: `1`)	`1`
`concat`	`bool`	If set to :obj:`False`, the multi-head-attentions are averaged instead of concatenated. (default: `True`)	`True`
`beta`	`bool`	If set to :obj:`True`, will use a gated residual connection. (default: `False`)	`False`
`dropout`	`float`	Dropout probability of the normalized attention coefficients. (default: `0.0`)	`0.0`
`edge_dim`	`Optional[int]`	Edge feature dimensionality (in case there are any). (default: :obj:`None`)	`None`
`bias`	`bool`	If set to :obj:`False`, the layer will not learn an additive bias. (default: `True`)	`True`
`root_weight`	`bool`	If set to :obj:`False`, the layer will not add the transformed root node features. (default: `True`)	`True`

GINConv

`k3_node.layers.conv.GINConv`

Bases: MessagePassing

The graph isomorphism operator from the "How Powerful are Graph Neural Networks?" <https://arxiv.org/abs/1810.00826>_ paper.

Parameters:

Name	Type	Description	Default
`nn`	`Union[Callable, int]`	A neural network :math:`h_{\mathbf{\Theta}}` that maps node features to new embeddings (e.g. a :class:`keras.Sequential` or callable). Also accepts an integer `channels` for backward compatibility.	required
`eps`	`float`	(Initial) :math:`\epsilon`-value. (default: `0.0`)	`0.0`
`train_eps`	`bool`	If set to :obj:`True`, :math:`\epsilon` will be a learnable parameter. (default: `False`)	`False`

GINEConv

`k3_node.layers.conv.GINEConv`

Bases: MessagePassing

The modified :class:GINConv operator from the "Strategies for Pre-training Graph Neural Networks" <https://arxiv.org/abs/1905.12265>_ paper, which is able to incorporate edge features into aggregation.

Parameters:

Name	Type	Description	Default
`nn`	`Callable`	A neural network :math:`h_{\mathbf{\Theta}}`.	required
`eps`	`float`	(Initial) :math:`\epsilon`-value. (default: `0.0`)	`0.0`
`train_eps`	`bool`	If set to :obj:`True`, :math:`\epsilon` will be a learnable parameter. (default: `False`)	`False`
`edge_dim`	`Optional[int]`	Edge feature dimensionality. (default: :obj:`None`)	`None`

ChebConv

`k3_node.layers.conv.ChebConv`

Bases: MessagePassing

The Chebyshev spectral graph convolutional operator from the "Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering" <https://arxiv.org/abs/1606.09375>_ paper.

Parameters:

Name	Type	Description	Default
`in_channels`	`int`	Size of each input sample.	required
`out_channels`	`int`	Size of each output sample.	required
`K`	`int`	Chebyshev filter size :math:`K`.	required
`normalization`	`Optional[str]`	The normalization scheme for the graph Laplacian (`"sym"`, `"rw"` or :obj:`None`). (default: `"sym"`)	`'sym'`
`bias`	`bool`	If set to :obj:`False`, the layer will not learn an additive bias. (default: `"True"`)	`True`

AGNNConv

`k3_node.layers.conv.AGNNConv`

Bases: MessagePassing

The graph attentional propagation layer from the "Attention-based Graph Neural Network for Semi-Supervised Learning" <https://arxiv.org/abs/1803.03735>_ paper.

TAGConv

`k3_node.layers.conv.TAGConv`

Bases: MessagePassing

The topology adaptive graph convolutional operator from the "Topology Adaptive Graph Convolutional Networks" <https://arxiv.org/abs/1710.10370>_ paper.

Parameters:

Name	Type	Description	Default
`in_channels`	`int`	Size of each input sample.	required
`out_channels`	`int`	Size of each output sample.	required
`K`	`int`	Number of hops :math:`K`. (default: `3`)	`3`
`bias`	`bool`	If set to :obj:`False`, the layer will not learn an additive bias. (default: `True`)	`True`
`normalize`	`bool`	Whether to apply symmetric normalization. (default: `True`)	`True`

ARMAConv

`k3_node.layers.conv.ARMAConv`

Bases: MessagePassing

The ARMA graph convolutional operator from the "Graph Neural Networks with Convolutional ARMA Filters" <https://arxiv.org/abs/1901.01343>_ paper.

SGConv

`k3_node.layers.conv.SGConv`

Bases: MessagePassing

The simple graph convolutional operator from the "Simplifying Graph Convolutional Networks" <https://arxiv.org/abs/1902.07153>_ paper.

Parameters:

Name	Type	Description	Default
`in_channels`	`int`	Size of each input sample.	required
`out_channels`	`int`	Size of each output sample.	required
`K`	`int`	Number of hops :math:`K`. (default: `1`)	`1`
`cached`	`bool`	If set to :obj:`True`, the layer will cache the computation of :math:`\mathbf{\hat{D}}^{-1/2} \mathbf{\hat{A}} \mathbf{\hat{D}}^{-1/2}`. (default: `False`)	`False`
`add_self_loops`	`bool`	If set to :obj:`False`, will not add self-loops. (default: `True`)	`True`
`bias`	`bool`	If set to :obj:`False`, the layer will not learn an additive bias. (default: `True`)	`True`

SSGConv

`k3_node.layers.conv.SSGConv`

Bases: MessagePassing

The simple spectral graph convolutional operator from the "Simple Spectral Graph Convolution" <https://arxiv.org/abs/2109.07191>_ paper.

Parameters:

Name	Type	Description	Default
`in_channels`	`int`	Size of each input sample.	required
`out_channels`	`int`	Size of each output sample.	required
`alpha`	`float`	Teleport probability :math:`\alpha`.	required
`K`	`int`	Number of hops :math:`K`. (default: `1`)	`1`
`cached`	`bool`	If set to :obj:`True`, the layer will cache normalization coefficients. (default: `False`)	`False`
`add_self_loops`	`bool`	If set to :obj:`False`, will not add self-loops. (default: `True`)	`True`
`bias`	`bool`	If set to :obj:`False`, the layer will not learn an additive bias. (default: `True`)	`True`

APPNP

`k3_node.layers.conv.APPNP`

Bases: MessagePassing

The approximate personalized propagation of neural predictions (APPNP) operator from the "Predict then Propagate: Combining Neural Networks with Personalized PageRank for Classification on Graphs" <https://arxiv.org/abs/1810.05997>_ paper.

Parameters:

Name	Type	Description	Default
`K`	`int`	Number of iterations :math:`K`.	required
`alpha`	`float`	Teleport probability :math:`\alpha`.	required
`dropout`	`float`	Dropout probability of edges or features during propagation. (default: `0.0`)	`0.0`
`cached`	`bool`	If set to :obj:`True`, the layer will cache the computation of normalization coefficients. (default: `False`)	`False`
`add_self_loops`	`bool`	If set to :obj:`False`, will not add self-loops. (default: `True`)	`True`
`normalize`	`bool`	Whether to apply symmetric normalization. (default: `True`)	`True`

APPNPConv

`k3_node.layers.conv.APPNPConv`

Bases: Conv

k3_node.layers.APPNPConv Implementation of Approximate Personalized Propagation of Neural Predictions

Parameters:

Name	Description	Default
`channels`	The number of output channels.	required
`alpha`	The teleport probability.	`0.2`
`propagations`	The number of propagation steps.	`1`
`mlp_hidden`	A list of hidden channels for the MLP.	`None`
`mlp_activation`	The activation function to use in the MLP.	`'relu'`
`dropout_rate`	The dropout rate for the MLP.	`0.0`
`activation`	The activation function to use in the layer.	`None`
`use_bias`	Whether to add a bias to the linear transformation.	`True`
`kernel_initializer`	Initializer for the `kernel` weights matrix.	`'glorot_uniform'`
`bias_initializer`	Initializer for the bias vector.	`'zeros'`
`kernel_regularizer`	Regularizer for the `kernel` weights matrix.	`None`
`bias_regularizer`	Regularizer for the bias vector.	`None`
`activity_regularizer`	Regularizer for the output.	`None`
`kernel_constraint`	Constraint for the `kernel` weights matrix.	`None`
`bias_constraint`	Constraint for the bias vector.	`None`
`**kwargs`	Additional keyword arguments.	`{}`

PNAConv

`k3_node.layers.conv.PNAConv`

Bases: MessagePassing

The Principal Neighbourhood Aggregation graph convolutional operator from the "Principal Neighbourhood Aggregation for Graph Nets" <https://arxiv.org/abs/2004.05718>_ paper.

GENConv

`k3_node.layers.conv.GENConv`

Bases: MessagePassing

The generalized graph convolution operator from the "DeeperGCN: All You Need to Train Deeper GCNs" <https://arxiv.org/abs/2006.07739>_ paper.

GatedGraphConv

`k3_node.layers.conv.GatedGraphConv`

Bases: MessagePassing

k3_node.layers.GatedGraphConv

Implementation of Gated Graph Convolution (GGC) layer

Parameters:

Name	Description	Default
`channels`	The number of output channels.	`None`
`n_layers`	The number of GGC layers to stack.	`None`
`activation`	Activation function to use.	`None`
`use_bias`	Whether to add a bias to the linear transformation.	`True`
`kernel_initializer`	Initializer for the `kernel` weights matrix.	`'glorot_uniform'`
`bias_initializer`	Initializer for the bias vector.	`'zeros'`
`kernel_regularizer`	Regularizer for the `kernel` weights matrix.	`None`
`bias_regularizer`	Regularizer for the bias vector.	`None`
`activity_regularizer`	Regularizer for the output.	`None`
`kernel_constraint`	Constraint for the `kernel` weights matrix.	`None`
`bias_constraint`	Constraint for the bias vector.	`None`
`**kwargs`	Additional arguments to pass to the `MessagePassing` superclass.	`{}`

ResGatedGraphConv

`k3_node.layers.conv.ResGatedGraphConv`

Bases: MessagePassing

The residual gated graph convolutional operator from the "Residual Gated Graph ConvNets" <https://arxiv.org/abs/1711.07553>_ paper.

Parameters:

Name	Type	Description	Default
`in_channels`	`Union[int, Tuple[int, int]]`	Size of each input sample, or a tuple for bipartite graphs.	required
`out_channels`	`int`	Size of each output sample.	required
`act`	`Union[str, Callable]`	Activation function :math:`\sigma` for gating. (default: `"sigmoid"`)	`'sigmoid'`
`edge_dim`	`Optional[int]`	Edge feature dimensionality. (default: :obj:`None`)	`None`
`root_weight`	`bool`	If set to :obj:`False`, will not add transformed root features. (default: `True`)	`True`
`bias`	`bool`	If set to :obj:`False`, will not learn an additive bias. (default: `True`)	`True`

SimpleConv

`k3_node.layers.conv.SimpleConv`

Bases: MessagePassing

A simple, parameter-free message passing operator.

Parameters:

Name	Type	Description	Default
`aggr`	`Union[str, List[str], Aggregation, None]`	The aggregation scheme to use (`"sum"`, `"mean"`, `"min"`, `"max"`, `"mul"`). (default: `"sum"`)	`'sum'`
`combine_root`	`Optional[str]`	The way to combine root node features with the aggregated output (`"sum"`, `"cat"`, `"self_loop"`, or :obj:`None`). (default: :obj:`None`)	`None`

GraphConv

`k3_node.layers.conv.GraphConv`

Bases: MessagePassing

The graph neural network operator from the "Weisfeiler and Leman Go Neural: Higher-order Graph Neural Networks" <https://arxiv.org/abs/1810.02244>_ paper.

Parameters:

Name	Type	Description	Default
`in_channels`	`Union[int, Tuple[int, int]]`	Size of each input sample, or a tuple for bipartite graphs.	required
`out_channels`	`int`	Size of each output sample.	required
`aggr`	`str`	The aggregation scheme to use (`"add"`, `"mean"`, `"max"`). (default: `"add"`)	`'add'`
`bias`	`bool`	If set to :obj:`False`, the layer will not learn an additive bias. (default: `"True"`)	`True`

MFConv

`k3_node.layers.conv.MFConv`

Bases: MessagePassing

The molecular fingerprint graph convolutional operator from the "Convolutional Networks on Graphs for Learning Molecular Fingerprints" <https://arxiv.org/abs/1509.09292>_ paper.

Parameters:

Name	Type	Description	Default
`in_channels`	`Union[int, Tuple[int, int]]`	Size of each input sample, or a tuple for bipartite graphs.	required
`out_channels`	`int`	Size of each output sample.	required
`max_degree`	`int`	The maximum degree of any node. (default: `10`)	`10`
`bias`	`bool`	If set to :obj:`False`, the layer will not learn an additive bias. (default: `True`)	`True`

Relational & Directed Convolutions

RGCNConv

`k3_node.layers.conv.RGCNConv`

Bases: MessagePassing

The relational graph convolutional operator from the "Modeling Relational Data with Graph Convolutional Networks" <https://arxiv.org/abs/1703.06103>_ paper.

FastRGCNConv

`k3_node.layers.conv.FastRGCNConv`

Bases: RGCNConv

See :class:RGCNConv.

RGATConv

`k3_node.layers.conv.RGATConv`

Bases: MessagePassing

The relational graph attentional operator from the "Relational Graph Attention Networks" <https://arxiv.org/abs/1904.05811>_ paper.

SignedConv

`k3_node.layers.conv.SignedConv`

Bases: MessagePassing

The signed graph convolutional operator from the "Signed Graph Convolutional Network" <https://arxiv.org/abs/1808.06354>_ paper.

DirGNNConv

`k3_node.layers.conv.DirGNNConv`

Bases: Layer

A directed graph neural network operator from the "Directed Graph Neural Networks" <https://arxiv.org/abs/2301.07663>_ paper.

AntiSymmetricConv

`k3_node.layers.conv.AntiSymmetricConv`

Bases: Layer

The anti-symmetric graph convolutional operator from the "Anti-Symmetric DGN: a Continuous approach to Deep Graph Neural Networks" <https://arxiv.org/abs/2202.13085>_ paper.

Advanced & Scaled Graph Convolutions

FiLMConv

`k3_node.layers.conv.FiLMConv`

Bases: MessagePassing

The FiLM graph convolutional operator from the "GNN-FiLM: Graph Neural Networks with Feature-wise Linear Modulation" <https://arxiv.org/abs/1906.12192>_ paper.

SuperGATConv

`k3_node.layers.conv.SuperGATConv`

Bases: MessagePassing

The self-supervised graph attentional operator from the "How to Find Your Friendly Neighborhood: Graph Attention Design with Self-Supervision" <https://openreview.net/forum?id=Wi5KUNlqWty>_ paper.

EGConv

`k3_node.layers.conv.EGConv`

Bases: MessagePassing

The Efficient Graph Convolution from the "Adaptive Filters and Aggregator Fusion for Efficient Graph Convolutions" <https://arxiv.org/abs/2104.01481>_ paper.

MixHopConv

`k3_node.layers.conv.MixHopConv`

Bases: MessagePassing

The MixHop graph convolutional operator from the "Higher-Order Graph Convolutional Networks via MixHop" <https://arxiv.org/abs/1905.00067>_ paper.

PDNConv

`k3_node.layers.conv.PDNConv`

Bases: MessagePassing

The pathfinder discovery network convolutional operator from the "Pathfinder Discovery Networks for Neural Message Passing" <https://arxiv.org/abs/2010.12878>_ paper.

FAConv

`k3_node.layers.conv.FAConv`

Bases: MessagePassing

The Frequency Adaptive Graph Convolution operator from the "Beyond Low-Frequency Information in Graph Convolutional Networks" <https://arxiv.org/abs/2101.00797>_ paper.

PANConv

`k3_node.layers.conv.PANConv`

Bases: MessagePassing

The path integral based convolution operator from the "Path Integral Based Convolution and Pooling for Graph Neural Networks" <https://arxiv.org/abs/2004.14805>_ paper.

LEConv

`k3_node.layers.conv.LEConv`

Bases: MessagePassing

The local extremum graph convolutional operator from the "ASAP: Adaptive Structure Aware Pooling for Learning Hierarchical Graph Representations" <https://arxiv.org/abs/1911.07979>_ paper.

Parameters:

Name	Type	Description	Default
`in_channels`	`Union[int, Tuple[int, int]]`	Size of each input sample, or a tuple for bipartite graphs.	required
`out_channels`	`int`	Size of each output sample.	required
`bias`	`bool`	If set to :obj:`False`, the layer will not learn an additive bias. (default: `True`)	`True`

ClusterGCNConv

`k3_node.layers.conv.ClusterGCNConv`

Bases: MessagePassing

The ClusterGCN graph convolutional operator from the "Cluster-GCN: An Efficient Algorithm for Training Deep and Large Graph Convolutional Networks" <https://arxiv.org/abs/1905.07953>_ paper.

Parameters:

Name	Type	Description	Default
`in_channels`	`int`	Size of each input sample.	required
`out_channels`	`int`	Size of each output sample.	required
`diag_lambda`	`float`	Diagonal enhancement coefficient :math:`\lambda`. (default: `0.0`)	`0.0`
`add_self_loops`	`bool`	If set to :obj:`False`, will not add self-loops. (default: `True`)	`True`
`bias`	`bool`	If set to :obj:`False`, the layer will not learn an additive bias. (default: `True`)	`True`

GCN2Conv

`k3_node.layers.conv.GCN2Conv`

Bases: MessagePassing

The graph convolutional operator from the "Simple and Deep Graph Convolutional Networks" <https://arxiv.org/abs/2007.02133>_ paper.

Parameters:

Name	Type	Description	Default
`channels`	`int`	Size of each input and output sample.	required
`alpha`	`float`	The strength of the initial residual connection :math:`\alpha`.	required
`theta`	`Optional[float]`	The hyperparameter for the identity mapping :math:`\theta`. (default: :obj:`None`)	`None`
`layer`	`Optional[int]`	The layer index :math:`l`. (default: :obj:`None`)	`None`
`shared_weights`	`bool`	If set to :obj:`True`, will use the same weights for :math:`\mathbf{X}` and :math:`\mathbf{X}_0`. (default: :obj:`True`)	`True`
`cached`	`bool`	If set to :obj:`True`, will cache the computation of normalization coefficients. (default: `False`)	`False`
`add_self_loops`	`bool`	If set to :obj:`False`, will not add self-loops. (default: `True`)	`True`
`normalize`	`bool`	Whether to apply symmetric normalization. (default: `True`)	`True`

LGConv

`k3_node.layers.conv.LGConv`

Bases: MessagePassing

The LightGCN operator from the "LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation" <https://arxiv.org/abs/2002.02126>_ paper.

Parameters:

Name	Type	Description	Default
`normalize`	`bool`	Whether to apply symmetric normalization. (default: `True`)	`True`

NNConv

`k3_node.layers.conv.NNConv`

Bases: MessagePassing

The continuous kernel-based convolutional operator from the "Neural Message Passing for Quantum Chemistry" <https://arxiv.org/abs/1704.01212>_ paper.

Parameters:

Name	Type	Description	Default
`in_channels`	`Union[int, Tuple[int, int]]`	Size of each input sample, or a tuple for bipartite graphs.	required
`out_channels`	`int`	Size of each output sample.	required
`nn`	`Callable`	A neural network :math:`h_{\mathbf{\Theta}}` that maps edge features to shape :obj:`[-1, in_channels * out_channels]`.	required
`aggr`	`str`	The aggregation scheme to use (`"add"`, `"mean"`, `"max"`). (default: `"add"`)	`'add'`
`root_weight`	`bool`	If set to :obj:`False`, the layer will not add the transformed root node features. (default: `True`)	`True`
`bias`	`bool`	If set to :obj:`False`, the layer will not learn an additive bias. (default: `True`)	`True`

CGConv

`k3_node.layers.conv.CGConv`

Bases: MessagePassing

The Crystal Graph Convolutional operator from the "Crystal Graph Convolutional Neural Networks for an Accurate and Interpretable Prediction of Material Properties" <https://arxiv.org/abs/1710.10324>_ paper.

Parameters:

Name	Type	Description	Default
`channels`	`Union[int, Tuple[int, int]]`	Size of each input sample, or a tuple for bipartite graphs.	required
`dim`	`int`	Edge feature dimensionality. (default: `0`)	`0`
`aggr`	`str`	The aggregation scheme to use (`"add"`, `"mean"`, `"max"`). (default: `"add"`)	`'add'`
`batch_norm`	`bool`	If set to :obj:`True`, will apply batch normalization. (default: `False`)	`False`
`bias`	`bool`	If set to :obj:`False`, the layer will not learn an additive bias. (default: `True`)	`True`

EdgeConv

`k3_node.layers.conv.EdgeConv`

Bases: MessagePassing

The edge convolutional operator from the "Dynamic Graph CNN for Learning on Point Clouds" <https://arxiv.org/abs/1801.07829>_ paper.

Parameters:

Name	Type	Description	Default
`nn`	`Callable`	A neural network :math:`h_{\mathbf{\Theta}}` that maps pair-wise node features to new edge representations.	required
`aggr`	`str`	The aggregation scheme to use (`"max"`, `"mean"`, `"sum"`). (default: `"max"`)	`'max'`

DynamicEdgeConv

`k3_node.layers.conv.DynamicEdgeConv`

Bases: EdgeConv

The dynamic edge convolutional operator from the "Dynamic Graph CNN for Learning on Point Clouds" <https://arxiv.org/abs/1801.07829>_ paper, which dynamically constructs a graph using :math:k-NN at each layer.

Parameters:

Name	Type	Description	Default
`nn`	`Callable`	A neural network :math:`h_{\mathbf{\Theta}}`.	required
`k`	`int`	Number of nearest neighbors. (default: `6`)	`6`
`aggr`	`str`	The aggregation scheme to use (`"max"`, `"mean"`, `"sum"`). (default: `"max"`)	`'max'`
`num_workers`	`int`	Number of workers (ignored in Keras backend).	`1`

GeneralConv

`k3_node.layers.conv.GeneralConv`

Bases: MessagePassing

A general GNN layer adapted from the "Design Space for Graph Neural Networks" <https://arxiv.org/abs/2011.08843>_ paper.

Heterogeneous & Hypergraph Convolutions

HeteroConv

`k3_node.layers.conv.HeteroConv`

Bases: Layer

A generic wrapper for computing graph convolution on heterogeneous graphs.

Parameters:

Name	Type	Description	Default
`convs`	`Dict[Tuple[str, str, str], Layer]`	A dictionary holding a bipartite GNN layer for each individual edge type.	required
`aggr`	`str`	The aggregation scheme to use for grouping node embeddings generated by different relations (:obj:`"sum"`, :obj:`"mean"`, :obj:`"min"`, :obj:`"max"`, :obj:`"cat"`, :obj:`None`). (default: :obj:`"sum"`)	`'sum'`

HGTConv

`k3_node.layers.conv.HGTConv`

Bases: MessagePassing

The Heterogeneous Graph Transformer (HGT) operator from the "Heterogeneous Graph Transformer" <https://arxiv.org/abs/2003.01332>_ paper.

Parameters:

Name	Type	Description	Default
`in_channels`	`int or Dict[str, int]`	Size of each input sample of every node type.	required
`out_channels`	`int`	Size of each output sample.	required
`metadata`	`Tuple[List[str], List[Tuple[str, str, str]]]`	Node types and edge types.	required
`heads`	`int`	Number of multi-head-attentions. (default: :obj:`1`)	`1`

HANConv

`k3_node.layers.conv.HANConv`

Bases: MessagePassing

The Heterogeneous Graph Attention Operator from the "Heterogeneous Graph Attention Network" <https://arxiv.org/abs/1903.07293>_ paper.

Parameters:

Name	Type	Description	Default
`in_channels`	`int or Dict[str, int]`	Size of each input sample of every node type.	required
`out_channels`	`int`	Size of each output sample.	required
`metadata`	`Tuple[List[str], List[Tuple[str, str, str]]]`	Node types and edge types.	required
`heads`	`int`	Number of multi-head-attentions. (default: :obj:`1`)	`1`
`negative_slope`	`float`	LeakyReLU angle of the negative slope. (default: :obj:`0.2`)	`0.2`

HEATConv

`k3_node.layers.conv.HEATConv`

Bases: MessagePassing

The heterogeneous edge-enhanced graph attentional operator from the "Heterogeneous Edge-Enhanced Graph Attention Network For Multi-Agent Trajectory Prediction" <https://arxiv.org/abs/2106.07161>_ paper.

Parameters:

Name	Type	Description	Default
`in_channels`	`int`	Size of each input sample.	required
`out_channels`	`int`	Size of each output sample.	required
`num_node_types`	`int`	The number of node types.	required
`num_edge_types`	`int`	The number of edge types.	required
`edge_type_emb_dim`	`int`	The embedding size of edge types.	required
`edge_dim`	`int`	Edge feature dimensionality.	required
`edge_attr_emb_dim`	`int`	The embedding size of edge features.	required
`heads`	`int`	Number of multi-head-attentions. (default: :obj:`1`)	`1`
`concat`	`bool`	Whether to concatenate multi-head attention. (default: :obj:`True`)	`True`
`negative_slope`	`float`	LeakyReLU angle. (default: :obj:`0.2`)	`0.2`
`root_weight`	`bool`	Whether to add root node features. (default: :obj:`True`)	`True`
`bias`	`bool`	Whether to learn an additive bias. (default: :obj:`True`)	`True`

HypergraphConv

`k3_node.layers.conv.HypergraphConv`

Bases: MessagePassing

The hypergraph convolutional operator from the "Hypergraph Convolution and Hypergraph Attention" <https://arxiv.org/abs/1901.08150>_ paper.

Parameters:

Name	Type	Description	Default
`in_channels`	`int`	Size of each input sample.	required
`out_channels`	`int`	Size of each output sample.	required
`use_attention`	`bool`	Whether to use hypergraph attention. (default: :obj:`False`)	`False`
`attention_mode`	`str`	Attention mode (:obj:`"node"` or :obj:`"edge"`). (default: :obj:`"node"`)	`'node'`
`heads`	`int`	Number of multi-head-attentions. (default: :obj:`1`)	`1`
`concat`	`bool`	Whether to concatenate heads. (default: :obj:`True`)	`True`
`negative_slope`	`float`	LeakyReLU angle. (default: :obj:`0.2`)	`0.2`
`bias`	`bool`	Whether to learn an additive bias. (default: :obj:`True`)	`True`

DNAConv

`k3_node.layers.conv.DNAConv`

Bases: MessagePassing

The dynamic neighborhood aggregation operator from the "Just Jump: Towards Dynamic Neighborhood Aggregation in Graph Neural Networks" <https://arxiv.org/abs/1904.04849>_ paper.

Parameters:

Name	Type	Description	Default
`channels`	`int`	Size of each input/output sample.	required
`heads`	`int`	Number of multi-head-attentions. (default: :obj:`1`)	`1`
`groups`	`int`	Number of groups for linear projections. (default: :obj:`1`)	`1`
`dropout`	`float`	Dropout probability. (default: :obj:`0.0`)	`0.0`
`cached`	`bool`	Whether to cache GCN normalization. (default: :obj:`False`)	`False`
`normalize`	`bool`	Whether to apply symmetric normalization. (default: :obj:`True`)	`True`
`add_self_loops`	`bool`	Whether to add self-loops. (default: :obj:`True`)	`True`
`bias`	`bool`	Whether to learn an additive bias. (default: :obj:`True`)	`True`

WLConv

`k3_node.layers.conv.WLConv`

Bases: Layer

The Weisfeiler Lehman (WL) operator from the "A Reduction of a Graph to a Canonical Form and an Algebra Arising During this Reduction" <https://www.iti.zcu.cz/wl2018/pdf/wl_paper_translation.pdf>_ paper.

Parameters:

Name	Type	Description	Default
`**kwargs`		Additional layer arguments.	`{}`

GPSConv

`k3_node.layers.conv.GPSConv`

Bases: Layer

The general, powerful, scalable (GPS) graph transformer layer from the "Recipe for a General, Powerful, Scalable Graph Transformer" <https://arxiv.org/abs/2205.12454>_ paper.

Parameters:

Name	Type	Description	Default
`channels`	`int`	Size of each input sample.	required
`conv`	`Layer`	The local message passing layer.	`None`
`heads`	`int`	Number of multi-head-attentions. (default: :obj:`1`)	`1`
`dropout`	`float`	Dropout probability. (default: :obj:`0.0`)	`0.0`
`act`	`str`	Activation function. (default: :obj:`"relu"`)	`'relu'`
`norm`	`str`	Normalization function. (default: :obj:`"batch_norm"`)	`'batch_norm'`

Point Cloud & Geometric Convolutions

PointNetConv

`k3_node.layers.conv.PointNetConv`

Bases: MessagePassing

The PointNet set abstraction layer from the "PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space" <https://arxiv.org/abs/1706.02413>_ paper.

PointConv

`k3_node.layers.conv.PointConv = PointNetConv` `module-attribute`

PointTransformerConv

`k3_node.layers.conv.PointTransformerConv`

Bases: MessagePassing

The Point Transformer layer from the "Point Transformer" <https://arxiv.org/abs/2012.09164>_ paper.

PointGNNConv

`k3_node.layers.conv.PointGNNConv`

Bases: MessagePassing

The PointGNN graph convolutional operator from the "Point-GNN: Graph Neural Network for 3D Object Detection in a Point Cloud" <https://arxiv.org/abs/2003.01251>_ paper.

PPFConv

`k3_node.layers.conv.PPFConv`

Bases: MessagePassing

The PPFNet graph convolutional operator from the "PPFNet: Global Context Aware Local Features for Robust 3D Point Matching" <https://arxiv.org/abs/1802.02669>_ paper.

FeaStConv

`k3_node.layers.conv.FeaStConv`

Bases: MessagePassing

The (fault-tolerant) feature-steered graph convolution operator from the "FeaStNet: Feature-Steered Graph Convolutions for 3D Shape Analysis" <https://arxiv.org/abs/1706.05206>_ paper.

GMMConv

`k3_node.layers.conv.GMMConv`

Bases: MessagePassing

The gaussian mixture model convolutional operator from the "Geometric Deep Learning on Graphs and Manifolds using Mixture Model CNNs" <https://arxiv.org/abs/1611.08402>_ paper.

GravNetConv

`k3_node.layers.conv.GravNetConv`

Bases: MessagePassing

The GravNet operator from the "Learning Representations of Irregular Particle-Detector Geometry with Distance-Weighted Graph Networks" <https://arxiv.org/abs/1902.07987>_ paper.

MeshCNNConv

`k3_node.layers.conv.MeshCNNConv`

Bases: MessagePassing

The MeshCNN convolutional operator from the "MeshCNN: A Network With An Edge" <https://arxiv.org/abs/1809.05910>_ paper.

Parameters:

Name	Type	Description	Default
`in_channels`	`int`	Size of each input sample.	required
`out_channels`	`int`	Size of each output sample.	required
`kernels`	`List[Layer]`	A list of 5 neural network layers that transform edge representations. (default: :obj:`None`)	`None`

XConv

`k3_node.layers.conv.XConv`

Bases: Layer

The convolutional operator on :math:\mathcal{X}-transformed points from the "PointCNN: Convolution On X-Transformed Points" <https://arxiv.org/abs/1801.07791>_ paper.

Parameters:

Name	Type	Description	Default
`in_channels`	`int`	Size of each input sample.	required
`out_channels`	`int`	Size of each output sample.	required
`dim`	`int`	Point cloud dimensionality.	required
`kernel_size`	`int`	Size of the convolving kernel.	required
`hidden_channels`	`int`	Dimensionality of lifted points.	`None`
`dilation`	`int`	Dilation factor. (default: :obj:`1`)	`1`
`bias`	`bool`	Whether to learn an additive bias. (default: :obj:`True`)	`True`
`num_workers`	`int`	Kept for PyG compatibility.	`1`

SplineConv

`k3_node.layers.conv.SplineConv`

Bases: MessagePassing

The spline-based convolutional operator from the "SplineCNN: Fast Geometric Deep Learning with Continuous B-Spline Kernels" <https://arxiv.org/abs/1711.08920>_ paper.

Parameters:

Name	Type	Description	Default
`in_channels`	`int or tuple`	Size of each input sample.	required
`out_channels`	`int`	Size of each output sample.	required
`dim`	`int`	Pseudo-coordinate dimensionality.	required
`kernel_size`	`int or List[int]`	Size of the convolving kernel.	required
`is_open_spline`	`bool or List[bool]`	If set to :obj:`False`, uses closed B-spline basis. (default: :obj:`True`)	`True`
`degree`	`int`	B-spline basis degree. (default: :obj:`1`)	`1`
`aggr`	`str`	The aggregation scheme to use (:obj:`"mean"`, :obj:`"add"`, :obj:`"max"`). (default: :obj:`"mean"`)	`'mean'`
`root_weight`	`bool`	Whether to add transformed root node features. (default: :obj:`True`)	`True`
`bias`	`bool`	Whether to learn an additive bias. (default: :obj:`True`)	`True`

Spektral-Compatible Convolutions

CrystalConv

`k3_node.layers.conv.CrystalConv`

Bases: MessagePassing

k3_node.layers.CrystalConv Implementation of Crystal Graph Convolutional Neural Networks (CGCNN) layer

Parameters:

Name	Description	Default
`aggregate`	Aggregation function to use (one of 'sum', 'mean', 'max').	`'sum'`
`activation`	Activation function to use.	`None`
`use_bias`	Whether to add a bias to the linear transformation.	`True`
`kernel_initializer`	Initializer for the `kernel` weights matrix.	`'glorot_uniform'`
`bias_initializer`	Initializer for the bias vector.	`'zeros'`
`kernel_regularizer`	Regularizer for the `kernel` weights matrix.	`None`
`bias_regularizer`	Regularizer for the bias vector.	`None`
`activity_regularizer`	Regularizer for the output.	`None`
`kernel_constraint`	Constraint for the `kernel` weights matrix.	`None`
`bias_constraint`	Constraint for the bias vector.	`None`
`**kwargs`	Additional arguments to pass to the `MessagePassing` superclass.	`{}`

DiffusionConv

`k3_node.layers.conv.DiffusionConv`

Bases: Conv

k3_node.layers.DiffusionConv Implementation of Diffusion Convolutional Neural Networks (DCNN) layer

Parameters:

Name	Description	Default
`channels`	The number of output channels.	required
`K`	The number of diffusion steps.	`6`
`activation`	Activation function to use.	`'tanh'`
`kernel_initializer`	Initializer for the `kernel` weights matrix.	`'glorot_uniform'`
`kernel_regularizer`	Regularizer for the `kernel` weights matrix.	`None`
`kernel_constraint`	Constraint for the `kernel` weights matrix.	`None`
`**kwargs`	Additional arguments to pass to the `Conv` superclass.	`{}`

GraphConvolution

`k3_node.layers.conv.GraphConvolution`

Bases: Layer

k3_node.layers.GraphConvolution Implementation of Graph Convolution (GCN) layer

Parameters:

Name	Description	Default
`units`	Positive integer, dimensionality of the output space.	required
`activation`	Activation function to use.	`None`
`use_bias`	Whether to add a bias to the linear transformation.	`True`
`final_layer`	Deprecated, use tf.gather or GatherIndices instead.	`None`
`input_dim`	Deprecated, use `keras.layers.Input` with `input_shape` instead.	`None`
`kernel_initializer`	Initializer for the `kernel` weights matrix.	`'glorot_uniform'`
`kernel_regularizer`	Regularizer for the `kernel` weights matrix.	`None`
`kernel_constraint`	Constraint for the `kernel` weights matrix.	`None`
`bias_initializer`	Initializer for the bias vector.	`'zeros'`
`bias_regularizer`	Regularizer for the bias vector.	`None`
`bias_constraint`	Constraint for the bias vector.	`None`
`**kwargs`	Additional arguments to pass to the `Layer` superclass.	`{}`

GraphAttention

`k3_node.layers.conv.GraphAttention`

Bases: Layer

k3_node.layers.GraphAttention Implementation of Graph Attention (GAT) layer

Parameters:

Name	Description	Default
`units`	Positive integer, dimensionality of the output space.	required
`attn_heads`	Positive integer, number of attention heads.	`1`
`attn_heads_reduction`	{'concat', 'average'} Method for reducing attention heads.	`'concat'`
`in_dropout_rate`	Dropout rate applied to the input (node features).	`0.0`
`attn_dropout_rate`	Dropout rate applied to attention coefficients.	`0.0`
`activation`	Activation function to use.	`'relu'`
`use_bias`	Whether to add a bias to the linear transformation.	`True`
`final_layer`	Deprecated, use tf.gather or GatherIndices instead.	`None`
`saliency_map_support`	Whether to support saliency map calculations.	`False`
`kernel_initializer`	Initializer for the `kernel` weights matrix.	`'glorot_uniform'`
`kernel_regularizer`	Regularizer for the `kernel` weights matrix.	`None`
`kernel_constraint`	Constraint for the `kernel` weights matrix.	`None`
`bias_initializer`	Initializer for the bias vector.	`'zeros'`
`bias_regularizer`	Regularizer for the bias vector.	`None`
`bias_constraint`	Constraint for the bias vector.	`None`
`attn_kernel_initializer`	Initializer for the attention kernel weights matrix.	`'glorot_uniform'`
`attn_kernel_regularizer`	Regularizer for the attention kernel weights matrix.	`None`
`attn_kernel_constraint`	Constraint for the attention kernel weights matrix.	`None`
`**kwargs`	Additional arguments to pass to the `Layer` superclass.	`{}`

PPNPPropagation

`k3_node.layers.conv.PPNPPropagation`

Bases: Layer

k3_node.layers.PPNPPropagation Implementation of PPNP layer

Parameters:

Name	Description	Default
`units`	Positive integer, dimensionality of the output space.	required
`final_layer`	Deprecated, use tf.gather or GatherIndices instead.	`None`
`input_dim`	Deprecated, use `keras.layers.Input` with `input_shape` instead.	`None`
`**kwargs`	Additional arguments to pass to the `Layer` superclass.	`{}`