eegproc.deep_learning.unsupervised.GNN.GCN_band_separated module

class eegproc.deep_learning.unsupervised.GNN.GCN_band_separated.BandSeparatedGCNEncoder(timesteps, t_down, n_channels=14, n_bands=3, gcn_units=(64, 32), temporal_pool_sizes=None, emb_dim=128, dropout=0.1, activation='relu', use_batch_norm=False, graph_self_loop_bias=2.0, graph_identity_mix=0.0, graph_adjacency_reg_weight=0.0001, name='encoder_band_separated_gcn', **kwargs)[source]

Bases: BaseEncoder

Band-separated residual GCN encoder for EEG sequence data.

Input shape:

(batch, timesteps, n_channels * n_bands)

The flattened feature axis is interpreted in channel-major order:

[ch0_band0, ch0_band1, …, ch1_band0, ch1_band1, …]

At each timestep, the input is reshaped to:

(batch, timesteps, n_channels, n_bands)

Each frequency band is then processed by its own independent GCN stack:

band_0 -> GCN stack 0 band_1 -> GCN stack 1 … band_k -> GCN stack k

Therefore, every band has its own:

  • GraphConv weights

  • learnable electrode adjacency matrices

  • normalization layers

  • residual projections

Bands never mix inside the graph convolutions. Their node-preserving graph outputs are concatenated only after all per-band GCN stacks have completed.

The output remains:

(batch, ceil(timesteps / t_down), emb_dim)

so it remains compatible with downstream BiLSTM sequence models.

Parameters:

  • timesteps (int)

  • t_down (int)

  • n_channels (int, default: 14)

  • n_bands (int, default: 3)

  • gcn_units (tuple[int, ...], default: (64, 32))

  • temporal_pool_sizes (tuple[int, ...] | None, default: None)

  • emb_dim (int, default: 128)

  • dropout (float, default: 0.1)

  • activation (str, default: 'relu')

  • use_batch_norm (bool, default: False)

  • graph_self_loop_bias (float, default: 2.0)

  • graph_identity_mix (float, default: 0.0)

  • graph_adjacency_reg_weight (float, default: 0.0001)

  • name (str, default: 'encoder_band_separated_gcn')

call(inputs, training=False)[source]

Encode each EEG frequency band with an independent electrode GCN.

Parameters:

training (bool, default: False)

get_adjacency_matrices()[source]

Return one set of learned adjacency matrices per frequency band.

Return type:

dict[str, dict[str, Tensor]]

get_band_features(inputs, training=False)[source]

Expose per-band post-GCN node-preserving sequences for diagnostics.

Parameters:

training (bool, default: False)

Return type:

dict[str, Tensor]

get_config()[source]

Return serializable configuration for the band-separated encoder.

Return type:

dict

property n_features: int

Flattened number of channel-band features per timestep.

class eegproc.deep_learning.unsupervised.GNN.GCN_band_separated.GCNDecoder(timesteps, n_channels, n_bands, t_down, gcn_units, temporal_pool_sizes, emb_dim=128, dropout=0.1, activation='relu', use_batch_norm=False, graph_self_loop_bias=2.0, graph_identity_mix=0.0, graph_adjacency_reg_weight=0.0001, name='decoder_gcn', **kwargs)[source]

Bases: Model

Graph-aware decoder for GCNEncoder latent sequences.

The decoder first restores temporal resolution. It then projects each latent timestep into a distinct feature vector for every electrode, reshapes that vector back to a graph, and applies residual graph-convolution blocks before reconstructing the frequency-band values at each node.

This is intentionally different from a purely temporal Conv1D decoder: the output is reconstructed as n_channels graph nodes rather than as one undifferentiated flat feature vector.

Parameters:

  • timesteps (int)

  • n_channels (int)

  • n_bands (int)

  • t_down (int)

  • gcn_units (tuple[int, ...])

  • temporal_pool_sizes (tuple[int, ...] | None)

  • emb_dim (int, default: 128)

  • dropout (float, default: 0.1)

  • activation (str, default: 'relu')

  • use_batch_norm (bool, default: False)

  • graph_self_loop_bias (float, default: 2.0)

  • graph_identity_mix (float, default: 0.0)

  • graph_adjacency_reg_weight (float, default: 0.0001)

  • name (str, default: 'decoder_gcn')

call(inputs, training=False)[source]

Decode latent sequences into channel-specific EEG graph signals.

Parameters:

training (bool, default: False)

compute_output_shape(input_shape)[source]

Return the decoder output shape.

fix_length(x)[source]

Trim or pad the reconstructed sequence to exactly timesteps.

Parameters:

x (Tensor)

Return type:

Tensor

classmethod from_encoder(encoder, name='decoder_gcn')[source]

Create a graph-aware decoder from a configured encoder.

Parameters:
Return type:

GCNDecoder

get_adjacency_matrices()[source]

Return normalized decoder electrode graphs for diagnostics.

Return type:

dict[str, Tensor]

get_config()[source]

Return serializable configuration for the decoder.

Return type:

dict

property n_features: int

Flattened number of reconstructed channel-band features.

eegproc.deep_learning.unsupervised.GNN.GCN_band_separated.GCNEncoder

alias of BandSeparatedGCNEncoder