neuraltrain.models.simplerconv.SimplerConv

class neuraltrain.models.simplerconv.SimplerConv(*, subject_layers_config: SubjectLayers | None = None, merger_config: ChannelMerger | None = None, n_hidden_channels: int = 512, kernel_sizes: tuple[int, ...] = (3, 2, 2, 2, 2), strides: tuple[int, ...] = (3, 2, 2, 2, 2), output_nonlin: bool = True, dropout: float = 0.0)[source][source]

Convolutional encoder inspired by BENDR / wav2vec 2.0.

Parameters:

  • subject_layers_config (SubjectLayers or None) – If set, prepend a per-subject linear projection.

  • merger_config (ChannelMerger or None) – If set, prepend a ChannelMerger for multi-montage support.

  • n_hidden_channels (int) – Number of hidden channels in each convolutional layer.

  • kernel_sizes (tuple of int) – Kernel size for each convolutional layer.

  • strides (tuple of int) – Stride for each convolutional layer. Must have the same length as kernel_sizes.

  • output_nonlin (bool) – Apply GELU after the last convolutional layer.

  • dropout (float) – Channel-dropout (Dropout1d) probability after each convolution.