MOE ADAPTERS

Arguments

Options

--virtual_bs_nint

Help: Virtual batch size iterations

• Default: 8

--clip_backbonestr

Help: Clip backbone

• Default: ViT-L/14

--prompt_templatestr

Help: Template string

• Default: a bad photo of a {}.

Implementation of MoE-Adapters from the CVPR 2024 paper “Boosting Continual Learning of Vision-Language Models via Mixture-of-Experts Adapters” Paper: https://arxiv.org/abs/2403.11549 Original code: https://github.com/JiazuoYu/MoE-Adapters4CL

Classes

class models.moe_adapters.MoEAdapters(backbone, loss, args, transform, dataset=None)

Bases: FutureModel

MoE Adapters – Boosting Continual Learning of Vision-Language Models via Mixture-of-Experts Adapters.

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

NAME: str = 'moe-adapters'

begin_task(dataset)

change_transform(dataset)

forward(x)

future_forward(x)

get_optimizer()

Return type:

Optimizer

get_parameters()

static get_parser(parser)

Return type:

ArgumentParser

net: Model

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

class models.moe_adapters.Model(args, dataset, device='cpu')

Bases: Module

forward(images, n_past_classes=0, n_seen_classees=None, train=False)

Return type:

Tensor