neuraltrain.models.fmri_mlp.FmriMlp

class neuraltrain.models.fmri_mlp.FmriMlp(*, hidden: int = 4096, n_blocks: int = 4, norm_type: str = 'ln', act_first: bool = False, n_repetition_times: int = 1, time_agg: Literal['in_mean', 'in_linear', 'out_mean', 'out_linear'] = 'out_linear', use_tr_embeds: bool = False, tr_embed_dim: int = 16, use_tr_layer: bool = False, out_dim: int | None = None, subject_layers_config: SubjectLayers | None = None, subject_layers_dim: Literal['input', 'hidden'] = 'hidden', output_head_config: Mlp | dict[str, Mlp] | None = None)[source][source]

Residual MLP for fMRI decoding, adapted from MindEye [1].

Parameters:

  • hidden (int) – Hidden dimension of the residual MLP blocks.

  • n_blocks (int) – Number of residual MLP blocks.

  • norm_type ({"ln", "bn"}) – Normalization layer type: "ln" for LayerNorm, "bn" for BatchNorm.

  • act_first (bool) – If True, apply activation before normalization in each block.

  • n_repetition_times (int) – Number of fMRI repetition times (TRs) expected in the input.

  • time_agg ({"in_mean", "in_linear", "out_mean", "out_linear"}) – Temporal aggregation strategy. "in_*" aggregates before the MLP; "out_*" aggregates after.

  • use_tr_embeds (bool) – Add learned per-TR embeddings to the input features.

  • tr_embed_dim (int) – Dimensionality of each TR embedding.

  • use_tr_layer (bool) – Use a grouped 1-D convolution (one group per TR) as the first projection instead of a shared nn.Linear.

  • out_dim (int or None) – Explicit output dimension. Overridden by n_outputs at build time.

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

  • subject_layers_dim ({"input", "hidden"}) – Dimension used for the subject-layer projection.

  • output_head_config (Mlp, dict[str, Mlp], or None) – Optional output head(s) appended after the MLP backbone.

References