Continual-Learning-with-Label-free-Replay-Buffer-and-Cluster-Preservation

Introduction

This repository provides code for class-incremental and domain-incremental continual learning experiments of proposed iSL-LRCP and iUL-LRCP methods in the paper.

The README is structured as follows:

• Feature Extraction: Instructions on extracting DINOv2 ViT-L/14 features for both class-incremental and domain-incremental datasets.
• Running Experiments: Guidelines on running continual learning experiments in both class-incremental and domain-incremental settings.

To reproduce results, firstly, DINOv2 features must be extracted and saved as .npy files for the according datasets, then experiments can be conducted on .npy files. This two-stage approach was faster and easier to run than single-stage runs.

Feature Extraction

Extracting DINOv2 Features for Class-Incremental Datasets

Overview

Extract_features_CI.py is a feature extraction script designed to extract DINOv2 ViT-L/14 features for class-incremental learning datasets. It supports multiple datasets and saves extracted features and labels in .npy format, which can be used in continual learning experiments.

Usage

Run the script with the following command:

python Extract_features_CI.py --dataset {dataset_name} --data_path ./data --output_path ./features

Supported Datasets

The script supports the following datasets, specified via the --dataset argument:

• CIFAR-100 (cifar100)
• Caltech256 (caltech256)
• Tiny ImageNet (tinyimagenet)
• ImageNet32 (imagenet32)

Dataset-Specific Requirements

Before running the script, ensure that:

• The dataset is downloaded and located in the specified --data_path directory.
• The --output_path directory exists or is writable, as extracted .npy files will be stored there.

ImageNet32 Special Parameters

For ImageNet32, you must specify either:

• --process_val → To extract validation features.
• --batch N → To extract training features from batch N (1-5).

Examples

Extracting features from batch 3 of ImageNet32 training data:

python Extract_features_CI.py --dataset imagenet32 --data_path ./data \
    --output_path ./features --batch 3

Extracting validation set features from ImageNet32:

python Extract_features_CI.py \
    --dataset imagenet32 --data_path ./data \ 
    --output_path ./features --process_val

Extracting DINOv2 Features for Domain-Incremental Datasets

CORe50 Dataset

Run the following command to extract features for the CORe50 dataset:

python Extract_features_core50.py --data_root ./data  --output_path ./features

R-MNIST Dataset

Run the following command to extract features for the R-MNIST dataset:

python Extract_features_r_mnist.py --data_root ./data --output_path ./features

Running Experiments

Class-Incremental Setting

Proposed and Baseline Methods

To run experiments in the class-incremental setting, use one of the following commands:

python {method}.py \
    --num_tasks num_tasks --classes_per_task classes_per_task \
    --train_features_path ./{dataset_name}_train_features.npy \
    --train_labels_path ./{dataset_name}_train_labels.npy \
    --test_features_path ./{dataset_name}_test_features.npy \
    --test_labels_path ./{dataset_name}_test_labels.npy

The {method} placeholder should be replaced with one of the following methods:

• iSL-LRCP (Supervised variant of our proposed method)
• iUL-LRCP (Unsupervised variant of our proposed method)
• ER-AML (Baseline method)
• iCaRL (Baseline method)

Offline Learning

To run an offline learning experiment, use the following command:

python offline_lft.py \
    --num_classes=num_classes \
    --train_features_path ./{dataset_name}_train_features.npy \
    --train_labels_path ./{dataset_name}_train_labels.npy \
    --test_features_path ./{dataset_name}_test_features.npy \
    --test_labels_path ./{dataset_name}_test_labels.npy

Domain-Incremental Setting

Proposed and Baseline Methods

To run experiments in the domain-incremental setting, use one of the following commands:

python {method}.py \
    --num_tasks num_tasks --classes_per_task classes_per_task \
    --train_features_path ./{dataset_name}_train_features.npy \
    --train_labels_path ./{dataset_name}_train_labels.npy \
    --test_features_path ./{dataset_name}_test_features.npy \
    --test_labels_path ./{dataset_name}_test_labels.npy

The {method} placeholder should be replaced with one of the following methods:

• iSL-LRCP_domain_inc (Supervised variant of our proposed method)
• ER-AML_domain_inc (Baseline method)
• iCaRL_domain_inc (Baseline method)

Offline Learning

To run an offline learning experiment, use the following command:

python offline_lft_domain_inc.py \
    --num_domains num_domains --classes_per_task classes_per_task \
    --train_features_path ./{dataset_name}_train_features.npy \
    --train_labels_path ./{dataset_name}_train_labels.npy \
    --test_features_path ./{dataset_name}_test_features.npy \
    --test_labels_path ./{dataset_name}_test_labels.npy