scvi.nn.FCLayers#

class scvi.nn.FCLayers(n_in, n_out, n_cat_list=None, n_cont=0, n_layers=1, n_hidden=128, dropout_rate=0.1, use_batch_norm=True, use_layer_norm=False, use_activation=True, bias=True, inject_covariates=True, activation_fn=<class 'torch.nn.modules.activation.ReLU'>, residual=False)[source]#

Bases: Module

A helper class to build fully-connected layers for a neural network.

Parameters:
n_in int

The dimensionality of the input

n_out int

The dimensionality of the output

n_cat_list Iterable[int] (default: None)

A list containing, for each category of interest, the number of categories. Each category will be included using a one-hot encoding.

n_cont int (default: 0)

The dimensionality of the continuous covariates

n_layers int (default: 1)

The number of fully-connected hidden layers

n_hidden int (default: 128)

The number of nodes per hidden layer

dropout_rate float (default: 0.1)

Dropout rate to apply to each of the hidden layers

use_batch_norm bool (default: True)

Whether to have BatchNorm layers or not

use_layer_norm bool (default: False)

Whether to have LayerNorm layers or not

use_activation bool (default: True)

Whether to have layer activation or not

bias bool (default: True)

Whether to learn bias in linear layers or not

inject_covariates bool (default: True)

Whether to inject covariates in each layer, or just the first (default).

activation_fn Module (default: <class 'torch.nn.modules.activation.ReLU'>)

Which activation function to use

residual bool (default: False)

Whether to add a residual skip connection around every block except the first, whenever that block’s input and output widths match. The first block is always excluded, so this requires n_layers >= 2 to have any effect.

Attributes table#

Methods table#

forward(x, *cat_list[, cont])

Forward computation on x.

inject_into_layer(layer_num)

Helper to determine if covariates should be injected.

set_online_update_hooks([hook_first_layer])

Set online update hooks.

Attributes#

FCLayers.training: bool#

Methods#

FCLayers.forward(x, *cat_list, cont=None, **kwargs)[source]#

Forward computation on x.

Parameters:
x Tensor

tensor of values with shape (n_in,)

cat_list int

list of category membership(s) for this sample

cont Tensor | None (default: None)

tensor of continuous covariates with shape (n_cont,)

kwargs

Extra per-call context threaded to the _apply_layer / _apply_batch_norm hooks for subclasses to consume; ignored by the base layers.

Returns:

torch.Tensor tensor of shape (n_out,)

FCLayers.inject_into_layer(layer_num)[source]#

Helper to determine if covariates should be injected.

Return type:

bool

FCLayers.set_online_update_hooks(hook_first_layer=True)[source]#

Set online update hooks.