univi.data.MultiModalDataset

class univi.data.MultiModalDataset(*args, **kwargs)

Bases: Dataset

Torch dataset over one AnnData per modality.

Each item is x_dict (modality name -> 1D float tensor), optionally followed by labels and/or reconstruction targets:

• x_dict

• (x_dict, y) when labels is given (y is a scalar tensor, or a dict of scalar tensors keyed by head name for multi-head labels)

• (x_dict, recon_targets) when recon_targets_spec is given, where recon_targets[mod] = {"successes": ..., "total_count": ...}

• (x_dict, y, recon_targets) when both are given

Use univi.data.collate_multimodal_xy_recon() (or univi.workflows.make_loader()) to batch items.

Parameters:

• adata_dict – {modality: AnnData}. With paired=True all objects must have identical, identically ordered obs_names.

• layer – Layer to read per modality (None reads .X); a string or a dict.

• X_key – "X" or an .obsm key per modality; a string or a dict.

• labels – Integer labels, or a dict of label arrays keyed by head name.

• modality_cfgs – Optional ModalityConfig list, used for categorical modalities read from .obs (input_kind="obs").

• recon_targets_spec – {modality: {"successes_layer": ..., "total_count_layer": ...}} for binomial / beta-binomial likelihoods.

__init__(adata_dict, layer=None, X_key='X', paired=True, device=None, labels=None, dtype=torch.float32, modality_cfgs=None, recon_targets_spec=None)

Methods

__init__(adata_dict[, layer, X_key, paired, ...])

Attributes

n_cells
obs_names