scglue.models.scglue.SCGLUETrainer

class scglue.models.scglue.SCGLUETrainer(net, lam_data=None, lam_kl=None, lam_graph=None, lam_align=None, lam_sup=None, dsc_steps=None, normalize_u=None, modality_weight=None, optim=None, lr=None, **kwargs)

Bases: GLUETrainer

Trainer for SCGLUE

Parameters:

• net (scglue.models.scglue.SCGLUE) – SCGLUE network to be trained

• lam_data (typing.Optional[float]) – Data weight

• lam_kl (typing.Optional[float]) – KL weight

• lam_graph (typing.Optional[float]) – Graph weight

• lam_align (typing.Optional[float]) – Adversarial alignment weight

• lam_sup (typing.Optional[float]) – Cell type supervision weight

• dsc_steps (typing.Optional[int]) – Number of discriminator steps per encoder-decoder step

• normalize_u (typing.Optional[bool]) – Whether to L2 normalize cell embeddings before decoder

• modality_weight (typing.Optional[typing.Mapping[str, float]]) – Relative modality weight (indexed by modality name)

• optim (typing.Optional[str]) – Optimizer

• lr (typing.Optional[float]) – Learning rate

• **kwargs – Additional keyword arguments are passed to the optimizer constructor

Methods

compute_losses	Compute loss functions
format_data	Format data tensors :rtype: typing.Tuple[typing.Mapping[str, torch.Tensor], typing.Mapping[str, torch.Tensor], typing.Mapping[str, torch.Tensor], typing.Mapping[str, torch.Tensor], typing.Mapping[str, torch.Tensor], typing.Mapping[str, torch.Tensor], torch.Tensor, torch.Tensor, torch.Tensor]
train_step	A single training step

Attributes

BURNIN_NOISE_EXAG
freeze_u	Whether to freeze cell embeddings
logger