Awesome-CIL

Awesome-CIL collects resources, reference implementations, and reproducible results for class-incremental learning (CIL), also known as continual, incremental, or lifelong learning.

For paper reading notes, see Paper Reading Notes.

🎉Surveys

Title	Venue	Year	Code
Class-Incremental Learning: A Survey	TPAMI	2024	Official
Continual Learning with Pre-Trained Models: A Survey	IJCAI	2024	Official
PyCIL: A Python Toolbox for Class-Incremental Learning	—	—	Official

🚀 Reproduced Methods

Title	Method	Venue	Year	Category	Reference implementation
iCaRL: Incremental Classifier and Representation Learning	iCaRL	CVPR	2017	Memory	Official
Co-Transport for Class-Incremental Learning	COIL	MM	2021	Classifier transport	—
DER: Dynamically Expandable Representation for Class Incremental Learning	DER	CVPR	2021	Backbone expansion	Official
DualPrompt: Complementary Prompting for Rehearsal-free Continual Learning	DualPrompt	ECCV	2022	Prompt	Official
FOSTER: Feature Boosting and Compression for Class-Incremental Learning	FOSTER	ECCV	2022	Backbone expansion	Official
Learning to Prompt for Continual Learning	L2P	CVPR	2022	Prompt	Official
CODA-Prompt: COntinual Decomposed Attention-based Prompting for Rehearsal-Free Continual Learning	CODA-Prompt	CVPR	2023	Prompt	Official
Hierarchical Decomposition of Prompt-Based Continual Learning: Rethinking Obscured Sub-optimality	HiDe-Prompt	NeurIPS	2023	Prompt	Official
A Unified Continual Learning Framework with General Parameter-Efficient Tuning	LAE	ICCV	2023	PEFT expansion	Official
Expandable Subspace Ensemble for Pre-Trained Model-Based Class-Incremental Learning	EASE	CVPR	2024	PEFT expansion	Official
InfLoRA: Interference-Free Low-Rank Adaptation for Continual Learning	InfLoRA	CVPR	2024	PEFT / LoRA	Official
Revisiting Class-Incremental Learning with Pre-Trained Models: Generalizability and Adaptivity are All You Need	APER / ADAM	IJCV	2024	PEFT / adapter	Official
Revisiting Class-Incremental Learning with Pre-Trained Models: Generalizability and Adaptivity are All You Need	SimpleCIL	IJCV	2024	PTM prototype	Official
BiLoRA: Almost-Orthogonal Parameter Spaces for Continual Learning	BiLoRA	CVPR	2025	PEFT / LoRA	Official
CL-LoRA: Continual Low-Rank Adaptation for Rehearsal-Free Class-Incremental Learning	CL-LoRA	CVPR	2025	PEFT / LoRA	Official
MOS: Model Surgery for Pre-Trained Model-Based Class-Incremental Learning	MOS	AAAI	2025	PEFT / adapter	Official
SD-LoRA: Scalable Decoupled Low-Rank Adaptation for Class Incremental Learning	SD-LoRA	ICLR	2025	PEFT / LoRA	Official

🌟 Getting Started

🕹️ Clone

git clone https://github.com/hongwei-zhao/Awesome-CIL.git
cd Awesome-CIL

🗂️ Dependencies

• torch==2.0.1+cu118
• torchvision==0.15.2+cu118
• timm==0.6.7
• numpy==1.26.3
• scipy==1.12.0
• scikit-learn==1.4.2
• Pillow==10.2.0
• PyYAML==6.0.1
• tqdm==4.66.2

🔑 Run an Experiment

Select a JSON file in configs/<method>/, set the dataset path, GPU IDs, class order seed, initial classes, increment, and method-specific hyperparameters, then run:

export CIL_DATA_ROOT=/path/to/Dataset
python main.py --config configs/ease/ease_cifar_B0_Inc10.json

Batch launchers are available under scripts/. For example:

bash scripts/train_ease.sh

The configuration directory and registered model_name must agree:

Configuration directory	model_name
configs/aper	adam_adapter
configs/bilora	bilora
configs/cllora	cllora
configs/coda_prompt	coda_prompt
configs/coil	coil
configs/der	der
configs/dualprompt	dualprompt
configs/ease	ease
configs/finetune	finetune
configs/foster	foster
configs/hidep	hidep
configs/icarl	icarl
configs/inflora	inflora
configs/l2p	l2p
configs/lae	lae
configs/mos	mos
configs/sdlora	sdlora
configs/simplecil	simplecil

Common configuration fields:

• model_name: Must match a model registered in utils/factory.py.
• dataset: One of cifar224, cub, imageneta, imagenetr, omnibenchmark, or vtab.
• data_path: Local dataset root. CIL_DATA_ROOT provides the default dataset root used by the data loader.
• device: GPU ID list, for example ["0"] or ["0", "1"].
• seed: Class-order seeds, for example [1993, 1994, 42].
• init_cls and increment: Initial and subsequent class counts. init_cls=0 starts directly with the incremental split.
• backbone_type: The PTM-CIL comparisons typically use ViT-B/16-IN21K (vit_base_patch16_224_in21k or pretrained_vit_b16_224_in21k).
• fixed_memory, memory_size, and memory_per_class: Exemplar-memory controls for methods that use rehearsal.

📚 Datasets

Dataset	Training examples	Test examples	Classes
CIFAR-100	50,000	10,000	100
CUB-200-2011	9,430	2,358	200
ImageNet-R	24,000	6,000	200
ImageNet-A	5,981	1,519	200
ObjectNet	26,509	6,628	200
OmniBenchmark	89,697	5,983	300
VTAB	1,796	8,619	50

☄️ Pre-trained Models

• ViT-B/16-IN1K — Pre-trained on ImageNet-21K and fine-tuned on ImageNet-1K; create it with timm.create_model("vit_base_patch16_224", pretrained=True).
• ViT-B/16-IN21K — Pre-trained on ImageNet-21K without ImageNet-1K fine-tuning; create it with timm.create_model("vit_base_patch16_224_in21k", pretrained=True).
• ViT-L/16-IN1K — Large Vision Transformer pre-trained on ImageNet-21K and fine-tuned on ImageNet-1K.
• ViT-B/16-DINO — Self-supervised Vision Transformer trained with DINO on ImageNet.
• ViT-B/16-SAM — Vision Transformer trained on SA-1B and fine-tuned for segmentation tasks such as COCO and ADE20K.
• ViT-B/16-MAE — Vision Transformer trained with masked autoencoding on ImageNet-21K.
• ViT-B/16-CLIP — Vision Transformer trained with CLIP on large-scale image-text data.
• ResNet-18 / ResNet-50 / ResNet-152 — ImageNet-1K pre-trained convolutional baselines.

📊 Reproduced Results

The result workbook is the single source of truth for reproduced metrics. It contains the following result sheets:

• seed1993 — CIFAR-100, CUB-200, ImageNet-R, OmniBenchmark, and VTAB with class-order seed 1993.
• seed1994 — The same benchmark suite with class-order seed 1994.
• seed42 — The same benchmark suite with class-order seed 42.
• ImageNet-R (INR) — ImageNet-R task-count ablations for 5, 10, 20, and 40 tasks.

The result-log mapping records the workbook cells, expected log names, parsed metrics, and verification status.

License

This repository is distributed under the existing MIT License. The original copyright notice is retained. Pre-trained weights, datasets, and incorporated third-party components remain subject to their respective licenses.

🤗 Acknowledgments

• LAMDA-PILOT
• Awesome-Incremental-Learning