deephyper.keras.layers.AttentionLinear#

class deephyper.keras.layers.AttentionLinear(*args: Any, **kwargs: Any)[source]#

Bases: Layer

Linear Attention.

The attention coefficient between node \(i\) and \(j\) is calculated as:

\[\text{tanh} (\textbf{a}_l\textbf{Wh}_i + \textbf{a}_r\textbf{Wh}_j)\]

where \(\textbf{a}_l\) and \(\textbf{a}_r\) are trainable vectors.

Parameters:

  • state_dim (int) – number of output channels.

  • attn_heads (int) – number of attention heads.

Methods

build

call

Apply the layer on input tensors.
__call__(*args: Any, **kwargs: Any) → Any#

Call self as a function.

call(inputs, **kwargs)[source]#

Apply the layer on input tensors.

Parameters:

inputs (list) – X (tensor): node feature tensor N (int): number of nodes targets (tensor): target node index tensor sources (tensor): source node index tensor degree (tensor): node degree sqrt tensor (for GCN attention)

Returns:

attention coefficient tensor

Return type:

attn_coef (tensor)