univi.refinement.RefinementConfig

class univi.refinement.RefinementConfig(max_epochs=1000, warmup_epochs=10, lr_head=0.0001, lr_encoder=1e-05, weight_decay_head=1e-06, weight_decay_encoder=1e-06, latent_weight=0.0, replay_weight=0.0, replay_epoch=10000, patience=50, min_delta=0.0, grad_clip=5.0, log_every=25)

Bases: object

Settings for head warmup followed by encoder/head fine-tuning.

max_epochs includes warmup. latent_weight anchors posterior means to an eval-mode copy of the original reference. replay_weight adds the original generative objective on a separate paired training loader.

__init__(max_epochs=1000, warmup_epochs=10, lr_head=0.0001, lr_encoder=1e-05, weight_decay_head=1e-06, weight_decay_encoder=1e-06, latent_weight=0.0, replay_weight=0.0, replay_epoch=10000, patience=50, min_delta=0.0, grad_clip=5.0, log_every=25)

Methods

__init__([max_epochs, warmup_epochs, ...])

Attributes

grad_clip
latent_weight
log_every
lr_encoder
lr_head
max_epochs
min_delta
patience
replay_epoch
replay_weight
warmup_epochs
weight_decay_encoder
weight_decay_head