braindecode.modules.FeedForwardBlock#

class braindecode.modules.FeedForwardBlock(emb_size, expansion, drop_p, activation=<class 'torch.nn.modules.activation.GELU'>, *, hidden_features=None, gated=False, bias=True, output_drop_p=0.0)[source]#

Feedforward network block.

Parameters:

  • emb_size (int) – Embedding dimension.

  • expansion (int) – Expansion factor for the hidden layer size.

  • drop_p (float) – Dropout probability.

  • activation (type[Module]) – Activation function constructor.

  • hidden_features (int | None) – Hidden width, overriding expansion * emb_size when supplied.

  • gated (bool) – Use a separate activated gate projection multiplied by the value projection, rather than activating the value projection directly. Gated blocks register named projections (fc1, fc_gate, fc2); ungated blocks retain their sequential numeric child names.

  • bias (bool) – Include a bias in each linear projection.

  • output_drop_p (float) – Dropout probability after the output projection. The ungated default does not add an output dropout module.

Examples

>>> import torch
>>> from braindecode.modules import FeedForwardBlock
>>> module = FeedForwardBlock(emb_size=32, expansion=2, drop_p=0.1)
>>> inputs = torch.randn(2, 10, 32)
>>> outputs = module(inputs)
>>> outputs.shape
torch.Size([2, 10, 32])

Methods

forward(x)[source]#

Runs the forward pass.

Return type:

Tensor