Getting Started

This guide provides step-by-step instructions for setting up CLArena, running your first continual learning experiment, and checking the results.

CLArena supports four machine learning paradigms:

• Continual Learning (CL): Learning new tasks sequentially while retaining previous knowledge
• Continual Unlearning (CUL): Selectively forgetting information in continual learning scenarios
  
• Multi-Task Learning (MTL): Learning multiple tasks simultaneously
• Single-Task Learning (STL): Traditional supervised learning for individual tasks

This quickstart focuses on the continual learning main experiment. For detailed documentation on other paradigms, see:

• Continual Learning (CL)
• Continual Unlearning (CUL)
• Multi-Task Learning (MTL)
• Single-Task Learning (STL)

1 Installation

Follow these steps to install CLArena in your Python environment:

# Create a new conda environment (optional but recommended)
conda create -n clarena-env python=3.12
conda activate clarena-env

• Option 1, install from PyPI (recommended):

pip install clarena

• Option 2, install from source:

# Clone the repository
git clone https://github.com/pengxiang-wang/continual-learning-arena
cd continual-learning-arena

pip install .

Important: GPU Configuration

CLArena installs the CPU version of PyTorch by default. For GPU acceleration, install the appropriate PyTorch version for your CUDA setup by following the official PyTorch installation guide.

2 Run Default Experiment

This section walks you through running a default continual learning main experiment to verify your installation and familiarize yourself with CLArena’s workflow.

CLArena uses configuration files to define experiment parameters. Download the example configurations:

Download Example Configs View on GitHub

1. Extract the downloaded zip file to your working directory
2. Rename the extracted folder to configs/

Run the following command to start training with the default configuration:

clarena train clmain experiment=example_clmain_train

This command will train a continual learning model using the default settings described below.

Default Experiment Settings (Click to expand)

The default experiment uses the following configuration:

Component	Setting	Description
CL Paradigm	Task-Incremental Learning (TIL)	Global seed: 1
CL Dataset	Permuted MNIST	Classification task with 10 tasks
Backbone Network	Multi-Layer Perceptron (MLP)	Hidden layers: 256, 100, 64 with batch normalization
CL Algorithm	Finetuning	Initializes from the last trained task
Training	SGD Optimizer	Learning rate: 0.001, 2 epochs per task, batch size: 64, CPU training

Customizing Experiments

To run experiments with different configurations, refer to the Configure CL Main Experiment documentation.

3 Check the Results

After executing the training command, CLArena automatically creates an output directory containing all experiment data and results.

Results are saved in outputs/example_clmain_train/[timestamp]/, where [timestamp] represents the execution time. This structure allows multiple experimental runs to be stored separately.

Experiment output directory structure

You can monitor experiment progress and check intermediate results at any time during execution. The output directory is updated in real-time as training progresses.

The types of results generated are controlled by configuration settings. For detailed information about, see the Output Results (CL) section.