braindecode.modules.ChannelMerger#

class braindecode.modules.ChannelMerger(out_channels=270, pos_dim=2048, dropout=0.2, invalid_value=-0.1)[source]#

Spatial Fourier-attention merge: n_chans -> out_channels.

Montage-agnostic spatial attention: each virtual channel softmax-attends over the input channels using a learned head weighted against a Fourier embedding of the electrode (x, y) positions. Matches neuraltrain.models.common.ChannelMergerModel (parity-gated); braindecode has no subjects, so per_subject=False and subject_ids is dropped.

Parameters:

  • out_channels (int) – Number of virtual output channels. The default is 270.

  • pos_dim (int) – Fourier embedding dimension. The default is 2048.

  • dropout (float) – Spatial-attention dropout radius (non-parametric, training only) – NOT a Bernoulli probability. Each training step draws a random center in normalized [0, 1]^2 position space and bans every channel within this radius from the softmax. Values >= 1 ban all channels and zero the output. The default is 0.2.

  • invalid_value (float) – Position value marking padded/invalid channels (masked out of the softmax). The default is -0.1.

Methods

forward(x, positions)[source]#

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

Return type:

Tensor