sispca.hcv_vi

Classes

SupervisionAnnData

Custom data class for supervision from an AnnData object.

HCV

Add HSIC loss to the latent representation in the VAE model.

HCVI

Training a HSIC-regulated VAE model with scvi-tools.

Module Contents

class sispca.hcv_vi.SupervisionAnnData(target_key, field_type, target_type, target_n_dim=1)

Custom data class for supervision from an AnnData object.

Parameters:

  • target_key (str) – key name of the target variable.

  • field_type (str) – field type of the AnnData object where the target variable is stored.

  • target_type (str) – ‘continuous’ or ‘categorical’. The type of the target variable.

  • target_n_dim (int) – The number of dimensions of the target variable. - If target_type == ‘categorical’, number of labels to predict. - If target_type == ‘continuous’, number of output dimensions in a regression model.

target_key
field_type
target_type
target_n_dim = 1
_sanity_check()
class sispca.hcv_vi.HCV(n_input: int, n_layers: int = 1, dropout_rate: float = 0.1, n_hidden: int = 128, gene_likelihood: Literal['zinb', 'nb', 'poisson', 'normal'] = 'normal', hsic_scale: float = 1.0, n_latent_sub=[10, 10], predict_target_from_latent_sub: bool = False, target_supervision_list: list[SupervisionAnnData] | None = None, **model_kwargs)

Bases: scvi.module.VAE

Add HSIC loss to the latent representation in the VAE model.

Parameters:

  • hsic_scale (float) – Scalar multiplier of the HSIC penalty.

  • n_latent_sub (list of int) – Number of latent dimensions for each subspace.

  • predict_target_from_latent_sub (bool) – Whether to predict target variables from the latent space.

  • target_supervision_list (list of SupervisionAnnData) – List of SupervisionAnnData objects if predict_target_from_latent_sub is True.

gene_likelihood = 'normal'
hsic_scale = 1.0
n_latent_sub = [10, 10]
n_subspace
predict_target_from_latent_sub = False
target_supervision_list = None
predict(x, use_posterior_mean=False)

Forward passing through the encoder and run prediction.

Parameters:

  • x (tensor) – input tensor

  • use_posterior_mean (bool) – whether to use the posterior mean of the latent distribution for prediction

prediction_loss(labelled_dataset: dict[str, torch.Tensor])

Calculate the mean squared error loss for the QC signal prediction.

generative(z, library, batch_index, cont_covs=None, cat_covs=None, size_factor=None, y=None, transform_batch=None)

Run the generative process.

loss(tensors, inference_outputs, generative_outputs, kl_weight: float = 1.0, labelled_tensors: dict[str, torch.Tensor] | None = None)
class sispca.hcv_vi.HCVI(adata: anndata.AnnData, n_hidden: int = 128, n_layers: int = 1, dropout_rate: float = 0.1, gene_likelihood: Literal['zinb', 'nb', 'poisson', 'normal'] = 'normal', hsic_scale: float = 1.0, n_latent_sub=[10, 10], predict_target_from_latent_sub: bool = False, target_supervision_list: list[SupervisionAnnData] | None = None, **model_kwargs)

Bases: scvi.model.SCVI

Training a HSIC-regulated VAE model with scvi-tools.

Parameters:

  • hsic_scale (float) – Scalar multiplier of the HSIC penalty.

  • n_latent_sub (list of int) – Number of latent dimensions for each subspace.

  • predict_target_from_latent_sub (bool) – Whether to predict target variables from the latent space.

  • target_supervision_list (list of SupervisionAnnData) – List of SupervisionAnnData objects if predict_target_from_latent_sub is True.

_module_cls
predict_target_from_latent_sub = False
n_latent_sub = [10, 10]
n_subspace
target_key_and_type = None
n_target
module
unsupervised_history_ = None
semisupervised_history_ = None
_model_summary_string = Multiline-String
Show Value
"""HCVI model with the following parameters:
------------------------------------------
*(architecture): n_layers = 1, dropout_rate = 0.1, gene_likelihood = normal
*(subspace) n_subspace = Uninferable, n_latent_sub = List.list(ctx=<Context.Load: 1>,
          elts=[ <Const.int l.353 at 0x7045541400a0>,
            <Const.int l.353 at 0x7045541400d0>]), hsic_scale = 1.0
*(supervision) predict_target = False, n_target = Uninferable"""
init_params_
was_pretrained = False
train(max_epochs: int | None = None, n_samples_per_label: float | None = None, check_val_every_n_epoch: int | None = None, train_size: float = 0.9, validation_size: float | None = None, shuffle_set_split: bool = True, batch_size: int = 128, accelerator: str = 'auto', devices: int | list[int] | str = 'auto', datasplitter_kwargs: dict | None = None, plan_kwargs: dict | None = None, **trainer_kwargs)

Trains the model using SCVI.

classmethod setup_anndata(adata: anndata.AnnData, layer: str | None = None, target_supervision_list: list[SupervisionAnnData] | None = None, batch_key: str | None = None, labels_key: str | None = None, size_factor_key: str | None = None, categorical_covariate_keys: list[str] | None = None, continuous_covariate_keys: list[str] | None = None, **kwargs)

%(summary)s.

Parameters:

  • %(param_adata)s

  • %(param_layer)s

  • %(param_batch_key)s

  • %(param_labels_key)s

  • %(param_size_factor_key)s

  • %(param_cat_cov_keys)s

  • %(param_cont_cov_keys)s