RANPAC

Arguments

Options

--rp_sizeint

Help: size of the random projection layer (L in the paper)

• Default: 10000

Slow Learner with Classifier Alignment.

Note

SLCA USES A CUSTOM BACKBONE (see feature_extractor_type argument)

param –feature_extractor_type:

the type of convnet to use. vit-b-p16 is the default: ViT-B/16 pretrained on Imagenet 21k (NO finetuning on ImageNet 1k)

Classes

class models.ranpac.RanPAC(backbone, loss, args, transform, dataset=None)

Bases: ContinualModel

RanPAC: Random Projections and Pre-trained Models for Continual Learning.

COMPATIBILITY: List[str] = ['class-il', 'domain-il', 'task-il']

NAME: str = 'ranpac'

begin_task(dataset)

end_task(dataset)

forward(x)

freeze_backbone(is_first_session=False)

get_parameters()

static get_parser(parser)

Return type:

ArgumentParser

get_scheduler()

net: RanPAC_Model

observe(inputs, labels, not_aug_inputs, epoch=0)

optimise_ridge_parameter(Features, Y)

replace_fc(trainloader)

setup_RP()