SPR

Arguments

Options

--spr_debug_mode0|1|True|False -> bool

Help: Run SPR with just a few iterations?

• Default: False

--delayed_buffer_sizeint

Help: Size of the delayed buffer.

• Default: 500

--fitting_lrfloat

Help: LR used during finetuining (classifier buffer fitting on P)

• Default: 0.002

--fitting_epochsint

Help: Number of epochs used during finetuining (classifier buffer fitting on P)

• Default: 50

--inner_train_epochsint

Help: Inner train epochs for SSL (base net)

• Default: 3000

--expert_train_epochsint

Help: Innert train epochs for SSL (expert)

• Default: 4000

--simclr_tempfloat

Help: Temperature for simclr SSL loss

• Default: 0.5

--fitting_sched_lr_stepsizeint

Help: Step size for the LR scheduler during finetuining (classifier buffer fitting on P)

• Default: 300

--fitting_sched_lr_gammafloat

Help: Gamma for the LR scheduler during finetuining (classifier buffer fitting on P)

• Default: 0.1

--fitting_batch_sizeint

Help: Batch size for finetuining (classifier buffer fitting on P)

• Default: 16

--fitting_clip_valuefloat

Help: Gradient clipping for finetuning

• Default: 0.5

--E_maxint

Help: Number of stochastic ensemble for expert

• Default: 5

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.spr.SimCLR(transform, temp=0.5, eps=1e-06, filter_bs_len=None, correlation_mask=None)

Bases: object

class models.spr.Spr(backbone, loss, args, transform, dataset=None)

Bases: ContinualModel

Implementation of Continual Learning on Noisy Data Streams via Self-Purified Replay from ICCV 2021.

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

NAME: str = 'spr'

OVERRIDE_SUPPORT_DISTRIBUTED = True

begin_task(dataset)

cluster_and_sample()

filter samples in delay buffer

end_task(dataset)

fit_buffer()

Fit finetuned model on purified buffer, before eval

forward(inputs)

static get_parser(parser)

observe(inputs, labels, not_aug_inputs, true_labels)

train_self_base()

Self Replay. train base model with samples from delay and purified buffer

train_self_expert()

Train expert model with samples from delay buffer only

Functions

models.spr.disable_linear(backbone)

models.spr.init_simclr_net(model, device=None)