lynx.model.HeteroAttnVGAE

class lynx.model.HeteroAttnVGAE(*args, **kwargs)

Bases: BaseModel

Learning latent manifold w/ Conditional VGAE on hetero-graph Generative path: DESI (u) -> Latent (z) -> Xenium (x)

Parameters:

• configs (ml_collections.ConfigDict)

• device (torch.device)

__init__(configs, device=torch.device)

Parameters:

• configs (ml_collections.ConfigDict)

• device (torch.device)

Methods

__init__(configs[, device])
checkpoint(curr_loss, min_loss, patience, ...)
evaluate(adata_ref, adata_query, graph_data)	Full model inference
fit(dataset, train_configs[, DEBUG, log_wandb])	Full model training
get_anneal_weight(beta, epoch, warmup_epochs)
guide(data)	Variational guide
load_state(save_path)
lognorm(x)
model(data)	Generative model
model_train(model, dataset, train_configs[, ...])
monitor_metrics(data, device[, key])	(Debug-only) Monitor latent factor correlations & reconstruction
plot_latent_corr(pz_corr_scores, qz_corr_scores)
plot_loss(train_losses, val_losses)
predict(data, device)	Get latent (z) & reconstructions from batched data object
set_desc(pbar, epoch, train_loss, val_loss)
setup(model, train_configs)	Setup optimizer & inference objects
train_step(model, dataloader, svi, device[, key])	Single-epoch training step
val_step(model, dataloader, svi, device[, key])	Single-epoch validation step

Attributes

init_model_weights
set_seed

evaluate(adata_ref, adata_query, graph_data, n_subgraphs=1, device=torch.device)

Full model inference

Parameters:

• adata_ref (scanpy.AnnData)

• adata_query (scanpy.AnnData)

• graph_data (HeteroDataset)

• n_subgraphs (int)

• device (torch.device)

fit(dataset, train_configs, DEBUG=False, log_wandb=False)

Full model training

guide(data)

Variational guide

model(data)

Generative model

predict(data, device)

Get latent (z) & reconstructions from batched data object