scvi.external.scbasset.ScBassetModule

class scvi.external.scbasset.ScBassetModule(n_cells, batch_ids=None, n_filters_init=288, n_repeat_blocks_tower=6, filters_mult=1.122, n_filters_pre_bottleneck=256, n_bottleneck_layer=32, batch_norm=True, dropout=0.0, l2_reg_cell_embedding=0.0)

Bases: BaseModuleClass

PyTorch implementation of ScBasset [Yuan and Kelley, 2022].

Original implementation in Keras: calico/scBasset

Parameters:

• n_cells (int) – Number of cells to predict region accessibility

• batch_ids (ndarray | None (default: None)) – Array of (n_cells,) with batch ids for each cell

• n_filters_init (int (default: 288)) – Number of filters for the initial conv layer

• n_repeat_blocks_tower (int (default: 6)) – Number of layers in the convolutional tower

• filters_mult (float (default: 1.122)) – Proportion by which the number of filters should increase in the convolutional tower

• n_bottleneck_layer (int (default: 32)) – Size of the bottleneck layer

• batch_norm (bool (default: True)) – Whether to apply batch norm across model layers

• dropout (float (default: 0.0)) – Dropout rate across layers, by default we do not do it for convolutional layers, but we do it for the dense layers

• l2_reg_cell_embedding (float (default: 0.0)) – L2 regularization for the cell embedding layer

Attributes table

training

Methods table

generative(region_embedding)	Generative method for the model.
inference(dna_code)	Inference method for the model.
loss(tensors, inference_outputs, ...)	Loss function for the model.

Attributes

ScBassetModule.training: bool

Methods

ScBassetModule.generative(region_embedding)

Generative method for the model.

Return type:

dict[str, Tensor]

ScBassetModule.inference(dna_code)

Inference method for the model.

Return type:

dict[str, Tensor]

ScBassetModule.loss(tensors, inference_outputs, generative_outputs)

Loss function for the model.

Return type:

LossOutput