sc_reconstruction.decoders.ReconTransformerDecoder

class sc_reconstruction.decoders.ReconTransformerDecoder(n_input, n_output, n_cat_list=None, dim_model=128, num_head=8, dim_hid=256, nlayers=6, dropout=0.1, distribution='normal', learning_rate=0.0001, reduce_lr_on_plateau=False, lr_factor=0.6, lr_patience=5, lr_threshold=0.001, lr_min=0.0, library_size_mode='none', weight_decay=0.0, use_adamw=False, **kwargs)

Bases: BaseReconstructionDecoder

Recon-style wrapper mirroring ReconMLPDecoder.

• self.module is a LightningModule.

• train(datamodule, **trainer_kwargs) — fit on a datamodule.

• decode(z_numpy, *cat_numpy) -> numpy — invert latent to expression.

• save / load — checkpoint round-trip.

Parameters:

• n_input (int)

• n_output (int)

• n_cat_list (Iterable[int] | None)

• dim_model (int)

• num_head (int)

• dim_hid (int)

• nlayers (int)

• dropout (float)

• distribution (str)

• learning_rate (float)

• reduce_lr_on_plateau (bool)

• lr_factor (float)

• lr_patience (int)

• lr_threshold (float)

• lr_min (float)

• library_size_mode (str)

• weight_decay (float)

• use_adamw (bool)

__init__(n_input, n_output, n_cat_list=None, dim_model=128, num_head=8, dim_hid=256, nlayers=6, dropout=0.1, distribution='normal', learning_rate=0.0001, reduce_lr_on_plateau=False, lr_factor=0.6, lr_patience=5, lr_threshold=0.001, lr_min=0.0, library_size_mode='none', weight_decay=0.0, use_adamw=False, **kwargs)

Parameters:

• n_input (int)

• n_output (int)

• n_cat_list (Iterable[int] | None)

• dim_model (int)

• num_head (int)

• dim_hid (int)

• nlayers (int)

• dropout (float)

• distribution (str)

• learning_rate (float)

• reduce_lr_on_plateau (bool)

• lr_factor (float)

• lr_patience (int)

• lr_threshold (float)

• lr_min (float)

• library_size_mode (str)

• weight_decay (float)

• use_adamw (bool)

Methods

__init__(n_input, n_output[, n_cat_list, ...])
decode(z, *cat_list[, decode_batch_size])	Decode the latent representation back.
load(path[, map_location])	Load weights into the existing module.
save(path)	Save the trained decoder.
train([datamodule])