eegproc.deep_learning.unsupervised.GNN.GraphConv module

Graph convolution layer used by EEGProc GCN encoders and decoders.

class eegproc.deep_learning.unsupervised.GNN.GraphConv.GraphConv(units, n_nodes, activation='relu', use_bias=True, self_loop_bias=2.0, identity_mix=0.0, adjacency_reg_weight=0.0001, **kwargs)[source]

Bases: Layer

Graph convolution with a shared, learnable electrode adjacency matrix.

The graph nodes are EEG electrodes. Any frequency bands are node features, so the same n_nodes x n_nodes electrode graph is applied to every band before the learned feature projection mixes the band/features dimension. Inputs may contain any number of leading dimensions, including the native sequence shape (batch, time, n_nodes, features); graph operations are broadcast over all leading dimensions without TimeDistributed.

The normalized adjacency is

A = (1 - identity_mix) * softmax(A_theta + self_loop_bias * I) `` + identity_mix * I``.

self_loop_bias makes training start from an electrode-preserving graph, while identity_mix can guarantee a permanent local contribution. A small entropy penalty discourages the adjacency rows from remaining close to uniform global averaging.

Parameters:

  • units (int) – Output feature dimensionality per electrode.

  • n_nodes (int) – Number of graph nodes/electrodes.

  • activation (str | None, default: 'relu') – Activation applied after graph aggregation and linear projection.

  • use_bias (bool, default: True) – Whether to add a feature bias.

  • self_loop_bias (float, default: 2.0) – Positive value added to the diagonal adjacency logits before softmax. With 14 electrodes, 2.0 starts with a diagonal weight around 0.36 before any optional identity mixture.

  • identity_mix (float, default: 0.0) – Fixed proportion of the identity matrix mixed into the learned graph. Must be in [0, 1]. The GCN already has residual blocks, so zero is the recommended initial setting.

  • adjacency_reg_weight (float, default: 0.0001) – Weight of the mean row-entropy penalty added through add_loss. Set to zero to disable adjacency regularization.

build(input_shape)[source]
call(X, training=False)[source]

Apply graph aggregation and a learned feature projection.

get_config()[source]

Returns the config of the object.

An object config is a Python dictionary (serializable) containing the information needed to re-instantiate it.

normalized_adjacency(dtype=None)[source]

Return the normalized electrode adjacency used by this layer.

Return type:

Tensor