scvi.external.gimvi.JVAE#

class scvi.external.gimvi.JVAE(dim_input_list, total_genes, indices_mappings, gene_likelihoods, model_library_bools, library_log_means, library_log_vars, n_latent=10, n_layers_encoder_individual=1, n_layers_encoder_shared=1, dim_hidden_encoder=64, n_layers_decoder_individual=0, n_layers_decoder_shared=0, dim_hidden_decoder_individual=64, dim_hidden_decoder_shared=64, dropout_rate_encoder=0.2, dropout_rate_decoder=0.2, n_batch=0, n_labels=0, dispersion='gene-batch', log_variational=True)[source]#

Bases: BaseModuleClass

Joint variational auto-encoder for imputing missing genes in spatial data.

Implementation of gimVI [Lopez et al., 2019].

Parameters:
dim_input_list list[int]

List of number of input genes for each dataset. If

the datasets have different sizes, the dataloader will loop on the smallest until it reaches the size of the longest one

total_genes int

Total number of different genes

indices_mappings list[ndarray | slice]

list of mapping the model inputs to the model output Eg: [[0,2], [0,1,3,2]] means the first dataset has 2 genes that will be reconstructed at location [0,2] the second dataset has 4 genes that will be reconstructed at [0,1,3,2]

gene_likelihoods list[str]

list of distributions to use in the generative process ‘zinb’, ‘nb’, ‘poisson’

model_library_bools list[bool]

model or not library size with a latent variable or use observed values

library_log_means list[ndarray | None]

List of 1 x n_batch array of means of the log library sizes. Parameterizes prior on library size if not using observed library sizes.

library_log_vars list[ndarray | None]

List of 1 x n_batch array of variances of the log library sizes. Parameterizes prior on library size if not using observed library sizes.

n_latent int (default: 10)

dimension of latent space

n_layers_encoder_individual int (default: 1)

number of individual layers in the encoder

n_layers_encoder_shared int (default: 1)

number of shared layers in the encoder

dim_hidden_encoder int (default: 64)

dimension of the hidden layers in the encoder

n_layers_decoder_individual int (default: 0)

number of layers that are conditionally batchnormed in the encoder

n_layers_decoder_shared int (default: 0)

number of shared layers in the decoder

dim_hidden_decoder_individual int (default: 64)

dimension of the individual hidden layers in the decoder

dim_hidden_decoder_shared int (default: 64)

dimension of the shared hidden layers in the decoder

dropout_rate_encoder float (default: 0.2)

dropout encoder

dropout_rate_decoder float (default: 0.2)

dropout decoder

n_batch int (default: 0)

total number of batches

n_labels int (default: 0)

total number of labels

dispersion str (default: 'gene-batch')

See vae.py

log_variational bool (default: True)

Log(data+1) prior to encoding for numerical stability. Not normalization.

Attributes table#

Methods table#

generative(z, library[, batch_index, y, ...])

Run the generative model.

get_sample_rate(x, batch_index, *_, **__)

Get the sample rate for the model.

inference(x[, mode, n_samples, batch_index])

Run the inference model.

loss(tensors, inference_outputs, ...[, ...])

Return the reconstruction loss and the Kullback divergences.

reconstruction_loss(x, px_rate, px_r, ...)

Compute the reconstruction loss.

sample_from_posterior_l(x[, mode, deterministic])

Sample the tensor of library sizes from the posterior.

sample_from_posterior_z(x[, mode, deterministic])

Sample tensor of latent values from the posterior.

sample_rate(x, mode, batch_index[, y, ...])

Returns the tensor of scaled frequencies of expression.

sample_scale(x, mode, batch_index[, y, ...])

Return the tensor of predicted frequencies of expression.

Attributes#

JVAE.training: bool#

Methods#

JVAE.generative(z, library, batch_index=None, y=None, mode=0, transform_batch=None)[source]#

Run the generative model.

Return type:

dict

JVAE.get_sample_rate(x, batch_index, *_, **__)[source]#

Get the sample rate for the model.

JVAE.inference(x, mode=0, n_samples=1, batch_index=None)[source]#

Run the inference model.

Return type:

dict

JVAE.loss(tensors, inference_outputs, generative_outputs, mode=None, kl_weight=1.0)[source]#

Return the reconstruction loss and the Kullback divergences.

Parameters:
x

tensor of values with shape (batch_size, n_input) or (batch_size, n_input_fish) depending on the mode

batch_index

array that indicates which batch the cells belong to with shape batch_size

y

tensor of cell-types labels with shape (batch_size, n_labels)

mode int | None (default: None)

indicates which head/tail to use in the joint network

Return type:

tuple[Tensor, Tensor]

Returns:

the reconstruction loss and the Kullback divergences

JVAE.reconstruction_loss(x, px_rate, px_r, px_dropout, mode)[source]#

Compute the reconstruction loss.

Return type:

Tensor

JVAE.sample_from_posterior_l(x, mode=None, deterministic=False)[source]#

Sample the tensor of library sizes from the posterior.

Parameters:
x Tensor

tensor of values with shape (batch_size, n_input) or (batch_size, n_input_fish) depending on the mode

mode int (default: None)

head id to use in the encoder

deterministic bool (default: False)

bool - whether to sample or not

Return type:

Tensor

Returns:

type tensor of shape (batch_size, 1)

JVAE.sample_from_posterior_z(x, mode=None, deterministic=False)[source]#

Sample tensor of latent values from the posterior.

Parameters:
x Tensor

tensor of values with shape (batch_size, n_input)

mode int (default: None)

head id to use in the encoder

deterministic bool (default: False)

bool - whether to sample or not

Return type:

Tensor

Returns:

type tensor of shape (batch_size, n_latent)

JVAE.sample_rate(x, mode, batch_index, y=None, deterministic=False, decode_mode=None)[source]#

Returns the tensor of scaled frequencies of expression.

Parameters:
x Tensor

tensor of values with shape (batch_size, n_input) or (batch_size, n_input_fish) depending on the mode

y Tensor | None (default: None)

tensor of cell-types labels with shape (batch_size, n_labels)

mode int

int encode mode (which input head to use in the model)

batch_index Tensor

array that indicates which batch the cells belong to with shape batch_size

deterministic bool (default: False)

bool - whether to sample or not

decode_mode int (default: None)

int use to a decode mode different from encoding mode

Return type:

Tensor

Returns:

type tensor of means of the scaled frequencies

JVAE.sample_scale(x, mode, batch_index, y=None, deterministic=False, decode_mode=None)[source]#

Return the tensor of predicted frequencies of expression.

Parameters:
x Tensor

tensor of values with shape (batch_size, n_input) or (batch_size, n_input_fish) depending on the mode

mode int

int encode mode (which input head to use in the model)

batch_index Tensor

array that indicates which batch the cells belong to with shape batch_size

y Tensor | None (default: None)

tensor of cell-types labels with shape (batch_size, n_labels)

deterministic bool (default: False)

bool - whether to sample or not

decode_mode int | None (default: None)

int use to a decode mode different from encoding mode

Return type:

Tensor

Returns:

type tensor of predicted expression