Continual Learning Main Experiment

Continual learning main experiment is the core experiment in CLArena for training and evaluating continual learning algorithms. This section defines its pipeline and guides you through configuring custom experiments.

Experiment Pipeline

For each task $t$ specified in the configuration:

1. Training Phase
   • The model is trained on current task data ${\mathcal{D}}_{\text{train}}^{(t)}$ for the specified number of epochs
   • The continual learning algorithm’s mechanisms are applied during this phase
   • Model parameters are updated while attempting to preserve knowledge from previous tasks
2. Validation Phase
   • At the end of each training epoch, the model is validated on ${\mathcal{D}}_{\text{val}}^{(t)}$
   • Validation results can optionally guide model selection and hyperparameter tuning
3. Evaluation Phase
   • After completing training and validation, the model is evaluated on test datasets
   • Testing occurs across all previously seen tasks: ${\mathcal{D}}_{\text{test}}^{(t_1)},\cdots ,{\mathcal{D}}_{\text{test}}^{(t)}$
   • The evaluation assesses both task-specific and overall performance
   • Note: This phase may be skipped based on configuration settings (see eval_after_tasks in required config fields)

Running

To run a continual learning main experiment, specify the CL_MAIN_EXPR indicator in the command:

clarena pipeline=CL_MAIN_EXPR index=<index-config-name>

Configuration

To run a custom continual learning main experiment, create a YAML file in the index/ folder as index config. Below is an example.

Example

configs/index/example_clmain_expr.yaml

# @package _global_
# make sure to include the above commented global setting!

# pipeline info
pipeline: CL_MAIN_EXPR
expr_name: example_clmain_expr
train_tasks: 10
eval_after_tasks: 10 
global_seed: 1

# paradigm settings
cl_paradigm: TIL

# components
defaults:
  - /cl_dataset: permuted_mnist.yaml
  - /backbone: clmlp.yaml
  - /cl_algorithm: finetuning.yaml
  - /optimizer: adam.yaml
  - /lr_scheduler: reduce_lr_on_plateau.yaml
  - /trainer: cpu.yaml
  - /metrics: cl_default.yaml
  - /lightning_loggers: default.yaml
  - /callbacks: cl_default.yaml
  - /hydra: default.yaml
  - /misc: default.yaml

# outputs
output_dir: outputs/${expr_name}/${misc.timestamp}

# overrides
trainer:
  max_epochs: 2

Required Config Fields

Below is the list of required config fields for the index config of continual learning main experiment.

Field	Description	Allowed Values
pipeline	The default pipeline that clarena use the config to run	Choose from supported pipeline indicators CL_MAIN_EXPR, CL_REF_JOINT_EXPR, CL_REF_INDEPENDENT_EXPR, CL_REF_RANDOM_EXPR, CL_FULL_EXPR
expr_name	The name of the experiment	A string value
train_tasks	The list of task IDs1 to train	List of integers $t_k$: At least 1, no more than the available number of tasks of CL dataset Integer $T$: Equivalent to list of integers [$1,\cdots ,T$]. At least 0 (no task to train), no more than the available number of tasks of CL dataset
eval_after_tasks	If task ID $t$ is in this list, run the evaluation process for all seen tasks after training task $t$	List of integers $t_k$ or integer $T$, works the same as train_tasks Must be a subset of train_tasks
global_seed	The global seed for the entire experiment	Same as the seed argument in lightning.seed_everything()
cl_paradigm	The continual learning paradigm	‘TIL’: Task-Incremental Learning (TIL) ‘CIL’: Class-Incremental Learning (CIL)
/cl_dataset	The continual learning dataset on which the experiment is conducted	Choose from sub-config YAML files in the cl_dataset/ folder See Configure CL Dataset
/cl_algorithm	The continual learning algorithm	Choose from sub-config YAML files in the cl_algorithm/ folder See Configure CL Algorithm
/backbone	The backbone network on which the continual learning algorithm is based	Choose from sub-config YAML files in the backbone/ folder See Configure Backbone Network
/optimizer	The optimizer for each task	Choose from sub-config YAML files in the optimizer/ folder See Configure Optimizer(s)
/lr_scheduler	The learning rate scheduler for each task	Choose from sub-config YAML files in the lr_scheduler/ folder See Configure Learning Rate Scheduler(s)
/trainer	The PyTorch Lightning Trainer object that contains all configurations for the training, validation and test process	Choose from sub-config YAML files in the trainer/ folder See Configure Trainer(s)
/metrics	The metrics to be monitored, logged, or visualized	Choose from sub-config YAML files in the metrics/ folder See Configure Metrics
/lightning_loggers	The Lightning Loggers used to log metrics and results	Choose from sub-config YAML files in the lightning_loggers/ folder See Configure Lightning Loggers
/callbacks	The callbacks applied to this experiment (other than metric callbacks). Callbacks are additional actions integrated at different points during the experiment	Choose from sub-config YAML files in the callbacks/ folder See Configure Callbacks
/hydra	Configuration for Hydra	Choose from sub-config YAML files in the hydra/ folder See Other Configs
/misc	Miscellaneous configurations that are less related to the experiment	Choose from sub-config YAML files in the misc/ folder See Other Configs
output_dir	The folder storing the experiment results	Relative path to where you run the clarena command We recommend including experiment names and timestamps in the path to distinguish multiple runs, e.g. outputs/{expr_name}/${misc.timestamp}

Note

The continual learning main experiment is managed by the CLMainExperiment class. To learn how these fields work, please refer to its source code.

Footnotes

1. The task IDs $t_k$ are integers ranging from 1 to the number of tasks in the CL dataset. Each corresponds to a task-specific dataset within the CL dataset.