LWS

Arguments

Options

--buf_lambda_logitsfloat

Help: Penalty weight BCE past logits.

• Default: 1

--kd_lambdafloat

Help: Penalty weight MSE clusters Logits (fixed to 1, not searched)

• Default: 1

--buf_lambda_clustersfloat

Help: Penalty weight BCE past clusters.

• Default: 1

--gammafloat

Help: Weight cluster contribution (Eq. 3 and 4)

• Default: 1

--kint

Help: Number of clusters

• Default: 8

--n_binint

Help: Number of bins

• Default: 4

--momentumfloat

Help: Momentum for weights update

• Default: 0.3

Rehearsal arguments

Arguments shared by all rehearsal-based methods.

--buffer_sizeint

Help: The size of the memory buffer.

• Default: None

--minibatch_sizeint

Help: The batch size of the memory buffer.

• Default: None

Classes

class models.lws.LwS(backbone, loss, args, transform, dataset=None)

Bases: ContinualModel

Implementation of “Towards Unbiased Continual Learning: Avoiding Forgetting in the Presence of Spurious Correlations”

COMPATIBILITY: List[str] = ['biased-class-il']

NAME: str = 'lws'

begin_epoch(epoch, dataset)

begin_task(dataset)

cluster_counts()

clustering(train_loader)

compute_stats()

end_task(dataset)

extract_features(train_loader)

forward(inputs)

freeze_classifiers(task_id, freeze_cluster=False)

get_classes_and_clusters(inputs)

get_initial_losses(dataset)

static get_parser(parser)

Return type:

ArgumentParser

get_task_weights()

get_weights(indexes)

init_classifiers()

observe(inputs, labels, not_aug_inputs, epoch, indexes)

update(target_losses, clusters_losses, indexes, epoch)

update_cluster_weights(indexes, epoch)