ICARL

Arguments

Options

--compute_theoretical_best0|1|True|False -> bool

Help: Compute NCM with the theoretical case where all the training samples are stored AND extend the buffer to equalize number of samples? (as in the original code)

• Default: 0

--use_original_icarl_transform0|1|True|False -> bool

Help: Use the original iCaRL transform?

• Default: 0

--opt_wdfloat

Help: Optimizer weight decay

• Default: 1e-05

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.icarl.ICarl(backbone, loss, args, transform, dataset=None)

Bases: ContinualModel

Continual Learning via iCaRL.

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

NAME: str = 'icarl'

begin_task(dataset)

static binary_cross_entropy(pred, y)

compute_class_means()

Computes a vector representing mean features for each class.

end_task(dataset)

forward(x)

get_loss(inputs, labels, task_idx, logits)

Computes the loss tensor.

Parameters:

• inputs (Tensor) – the images to be fed to the network

• labels (Tensor) – the ground-truth labels

• task_idx (int) – the task index

• logits (Tensor) – the logits of the old network

Returns:

the differentiable loss value

Return type:

Tensor

static get_parser(parser)

Return type:

ArgumentParser

observe(inputs, labels, not_aug_inputs, logits=None, epoch=None)

wd()

Functions

models.icarl.c100_transform(inputs)

Original augmentation from https://github.com/srebuffi/iCaRL/blob/master/iCaRL-TheanoLasagne/utils_cifar100.py