JVAETrainer

class scvi.inference.JVAETrainer(model, discriminator, gene_dataset_list, train_size=0.9, use_cuda=True, kappa=1.0, n_epochs_kl_warmup=400, **kwargs)

Bases: scvi.inference.trainer.Trainer

The trainer class for the unsupervised training of JVAE.

Parameters

• model (ModuleModule) – A model instance from class JVAE

• discriminator (ModuleModule) – A model instance of a classifier (with logit output)

• gene_dataset_list (List[GeneExpressionDataset]List[GeneExpressionDataset]) – list of gene_dataset instance like [CortexDataset(), SmfishDataset()]

• train_size (floatfloat) – Train-test split ratio in (0,1) to split cells

• kappa (floatfloat) – float to weight the discriminator loss

• n_epochs_kl_warmup (intint) – Number of epochs for linear warmup of KL(q(z|x)||p(z)) term. After n_epochs_kl_warmup, the training objective is the ELBO. This might be used to prevent inactivity of latent units, and/or to improve clustering of latent space, as a long warmup turns the model into something more of an autoencoder.

• **kwargs – Other keywords arguments from the general Trainer class.

Attributes Summary

default_metrics_to_monitor
posteriors_loop

Methods Summary

data_loaders_loop()	returns an zipped iterable corresponding to loss signature
get_discriminator_confusion()	A good mixing should lead to a uniform matrix.
get_imputed_values([deterministic, …])	Return imputed values for all genes for each dataset
get_latent([deterministic])	Return the latent space embedding for each dataset
get_loss_magnitude([one_sample])	Return the different losses of the model separately.
loss(tensors[, return_details])	Compute the loss of vae (reconstruction + kl_divergence)
loss_discriminator(latent_tensors[, …])	Compute the loss of the discriminator (either for the true labels or the fool labels)
on_epoch_begin()
on_training_loop(tensors_list)
training_extras_end()	Place to put extra models in eval mode, etc.
training_extras_init([lr_d, eps])	Other necessary models to simultaneously train

Attributes Documentation

default_metrics_to_monitor = ['elbo']

posteriors_loop

Methods Documentation

data_loaders_loop()

returns an zipped iterable corresponding to loss signature

get_discriminator_confusion()

A good mixing should lead to a uniform matrix.

Return type

ndarrayndarray

get_imputed_values(deterministic=True, normalized=True, decode_mode=None)

Return imputed values for all genes for each dataset

Parameters

• deterministic (boolbool) – If true, use the mean of the encoder instead of a Gaussian sample for the latent vector

• normalized (boolbool) – Return imputed normalized values or not

• decode_mode (int, NoneOptional[int]) – If a decode_mode is given, use the encoder specific to each dataset as usual but use the decoder of the dataset of id decode_mode to impute values

Return type

List[ndarray]List[ndarray]

get_latent(deterministic=True)

Return the latent space embedding for each dataset

Parameters

deterministic (boolbool) – If true, use the mean of the encoder instead of a Gaussian sample

Return type

List[ndarray]List[ndarray]

get_loss_magnitude(one_sample=False)

Return the different losses of the model separately. Useful to inspect and compare their magnitude.

Parameters

one_sample (boolbool) – Use only one batch to estimate the loss, can be much faster/less exact on big datasets

Return type

Tuple[ndarray, ndarray, ndarray]Tuple[ndarray, ndarray, ndarray]

loss(tensors, return_details=False)

Compute the loss of vae (reconstruction + kl_divergence)

Parameters

• tensors (Iterable[Tensor]Iterable[Tensor]) – Tensors of observations for each dataset

• return_details (boolbool) – Boolean used to inspect the loss values, return detailed loss for each dataset

Return type

Tensor, Tuple[List[Tensor], List[Tensor]]Union[Tensor, Tuple[List[Tensor], List[Tensor]]]

Returns

type scalar loss if return_details is False, else tuple (reconstruction_loss, kl_loss)

loss_discriminator(latent_tensors, predict_true_class=True, return_details=False)

Compute the loss of the discriminator (either for the true labels or the fool labels)

Parameters

• latent_tensors (List[torch.Tensor]) – Tensors for each dataset of the latent space

• predict_true_class (boolbool) – Specify if the loss aims at minimizing the accuracy or the mixing

• return_details (boolbool) – Boolean used to inspect the loss values, return detailed loss for each dataset

• latent_tensors –

Return type

List[Tensor], TensorUnion[List[Tensor], Tensor]

Returns

type scalar loss if return_details is False, else list of scalar losses for each dataset

on_epoch_begin()

on_training_loop(tensors_list)

training_extras_end()

Place to put extra models in eval mode, etc.

training_extras_init(lr_d=0.001, eps=0.01)

Other necessary models to simultaneously train

Parameters

**extras_kwargs –

Returns