ZSCL

Arguments

Options

--clip_backbonestr

Help: Clip backbone

• Default: ViT-L/14

• Choices: ViT-B/16, ViT-L/14

--prompt_templatestr

Help: Template string

• Default: a good photo of a {}.

--weint

Help: Whether to use weight averaging

• Default: 1

--avg_freqint

Help: Frequency of weight averaging

• Default: 100

--lsfloat

Help: Label smoothing

• Default: 0.2

Implementation of ZSCL from the ICCV 2023 paper “Preventing zero-shot transfer degradation in continual learning of vision-language models” Paper: https://arxiv.org/abs/2303.06628 Original code: https://github.com/Thunderbeee/ZSCL

Classes

class models.zscl.ZSCL(backbone, loss, args, transform, dataset=None)

Bases: FutureModel

ZSCL – Preventing zero-shot transfer degradation in continual learning of vision-language models.

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

NAME: str = 'zscl'

begin_task(dataset)

change_transform(dataset)

distillation(t, s, T=2)

end_task(dataset)

forward(image)

future_forward(image)

get_optimizer()

get_parameters()

static get_parser(parser)

Return type:

ArgumentParser

merge_we(model_0, model_1, sma_count)

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