geeks-z/cil

CIL（Class-Incremental Learning）类增量学习【Life-Long Machine Learning/Continual Learning】

Awesome-CIL

🎉 介绍

CIL: Class-Incremental Learning/类增量学习【Life-Long Learning/Continual Learning】

🚀 Survey

Title	Venue	Year	Code	Code_Position
A survey on few-shot class-incremental learning	Neural Networks	2024
Few-shot Class-incremental Learning: A Survey

🌟 Papers

Title	Venue	Year	Type	Code	Code_Position
Expandable Subspace Ensemble for Pre-Trained Model-Based Class-Incremental Learning	CVPR	2024	Adapter	Official	📁 CVPR24-Ease-main
InfLoRA: Interference-Free Low-Rank Adaptation for Continual Learning	CVPR	2024	LoRA	Official	📁 InfLoRA-main
DualPrompt: Complementary Prompting for Rehearsal-free Continual Learning	ECCV	2022	Prompt	Official	📁 dualprompt-pytorch-master
FOSTER: Feature Boosting and Compression for Class-Incremental Learning	ECCV	2022	Prompt	Official	📁 ECCV22-FOSTER-master
PILOT: A Pre-Trained Model-Based Continual Learning Toolbox			Toolbox	Official	📁 LAMDA-PILOT-main
Learning to Prompt for Continual Learning	CVPR	2022		Official	📁 Learn-to-prompt-for-Continual-Learning-main
PyCIL: A Python Toolbox for Class-Incremental Learning			Toolbox	Official	📁 PyCIL-master
Revisiting Class-Incremental Learning with Pre-Trained Models: Generalizability and Adaptivity are All You Need				Official	📁 RevisitingCIL-main

📝 Reproduced Results

• `B-$m$ Inc-$n$' stands for the class split. $m$ represents the number of categories in the initial incremental task, while $n$ denotes the number of subsequent incremental tasks, with categories in these tasks evenly distributed. If $m = 0$, all categories in the dataset are evenly distributed across $n$ incremental tasks.
• ViT-B/16-IN21K as the pre-trained backbone
• 实验日志位于每个方法文件夹下的res目录
• results：CNN/NME

CIFAR-100

PEFT

	B0 Inc5	B0 Inc10	B0 Inc20	B50 Inc5	B0 Inc10	Code
Ease
InfLoRA
CODA-Prompt
DualPrompt
L2P

Memory

	B0 Inc5	B0 Inc10	B0 Inc20	B50 Inc5	B0 Inc10	Code
FOSTER						📁 LAMDA-PILOT-main
DER						📁 LAMDA-PILOT-main
iCaRL	87.34/90.9	85.74/90.22	84.07/89.43	83.46/88.83	80.67/87.06	📁 LAMDA-PILOT-main

CUB-200

ImageNet-A

Omnibenchmark

VTAB

👨‍🏫 Acknowledgments