spd_learn.models.GyroAtt#

class spd_learn.models.GyroAtt(n_patches: int = 6, n_chans: int = 22, n_outputs: int = 2, temporal_out_channels: int = 20, temporal_kernel_size: int = 25, temporal_padding: int | None = None, attention_in_features: int = 20, attention_out_features: int = 18, attention_type: str = 'gyro', metric: str = 'LEM', drop_prob: float = 0.0, covariance_method=<function sample_covariance>, spatial_filters: int | None = None, att_out_features: int | None = None, temporal_filters: int | None = None)[source]#

Bases: MAtt

Gyrovector Riemannian Attention Network for EEG Decoding.

Implements the full deep architecture for EEG decoding introduced in GyroAtt [Wang et al., 2025].

The architecture comprises:

  1. Spatiotemporal Frontend: Spatial and temporal 2D convolutions (inherited from MAtt).

  2. Patch Covariance Embedding: Signal segmentation into patches via PatchEmbeddingLayer, sample covariance estimation via CovLayer, and trace normalization via TraceNorm.

  3. Riemannian Attention: Either GyroSPDAttention (orthogonal gyro-automorphisms and Fréchet aggregation) or SymSPDAttention (tangent space projections via SPDLinear).

  4. Classification Backend: Riemannian ReEig activation, LogEig tangent projection, dropout, and linear classification.

Parameters:

  • n_patches (int, default=6) – Number of temporal patches.

  • n_chans (int, default=22) – Number of EEG channels.

  • n_outputs (int, default=2) – Number of output classification classes.

  • temporal_out_channels (int, default=20) – Number of output channels for the temporal convolution.

  • temporal_kernel_size (int, default=25) – Kernel size for the temporal convolution along the time dimension.

  • temporal_padding (int, optional) – Padding along time for the temporal convolution. Defaults to temporal_kernel_size // 2.

  • attention_in_features (int, default=20) – Input feature dimension for the attention module.

  • attention_out_features (int, default=18) – Feature dimensionality of the attention output SPD matrices.

  • attention_type ({"gyro", "sym"}, default="gyro") – Type of Riemannian attention module to use.

  • metric ({"LEM", "LCM"}, default="LEM") – Riemannian metric used in the attention mechanism.

  • drop_prob (float, default=0.0) – Dropout probability before the final linear classifier.

  • covariance_method (callable, default=sample_covariance) – Method used for computing sample covariances.

  • spatial_filters (int, optional) – Alias for temporal_out_channels and attention_in_features.

  • att_out_features (int, optional) – Alias for attention_out_features.

  • temporal_filters (int, optional) – Alias for intermediate feature dimension.

References

[1]

Rui Wang, Chen Hu, Xiaoning Song, Xiaojun Wu, Nicu Sebe, and Ziheng Chen. Towards a general attention framework on gyrovector spaces for matrix manifolds. In Advances in Neural Information Processing Systems, volume 38. 2025. URL: https://openreview.net/forum?id=ChenHu-ML/GyroAtt.

Examples

>>> import torch
>>> from spd_learn.models import GyroAtt
>>> model = GyroAtt(n_chans=22, n_outputs=2, attention_type="gyro")
>>> x = torch.randn(4, 22, 1000)
>>> out = model(x)
>>> out.shape
torch.Size([4, 2])
forward(x: Tensor) → Tensor[source]#

Forward pass of GyroAtt.

Parameters:

x (torch.Tensor) – Input tensor of shape (batch, n_chans, n_times) or precomputed covariance patches of shape (batch, n_patches, spatial_filters, spatial_filters).

Returns:

Logits of shape (batch, n_outputs).

Return type:

torch.Tensor