braindecode.modules.PreNormTransformerBlock

class braindecode.modules.PreNormTransformerBlock(dim, attn, ffn, drop_prob)

Pre-norm transformer block as a plain Sequential of two residuals.

Residual(LayerNorm -> attn -> Dropout) then Residual(LayerNorm -> ffn -> Dropout); the children are indexed (0.fn.0, 0.fn.1, …), as in EEGConformer’s released layout. EEGConformer and MIRepNet use it.

Parameters:

• dim (int) – Token width, normalized before each sub-layer.

• attn (Module) – Token-mixing sub-layer, (batch, seq, dim) -> (batch, seq, dim).

• ffn (Module) – Feed-forward sub-layer, same shape.

• drop_prob (float) – Dropout after each sub-layer.