/TSCIL

[KDD2024] Class-incremental Learning for Time Series: Benchmark and Evaluation

Primary LanguagePython

Class-incremental Learning for Time Series: Benchmark and Evaluation

A unified experimental framework for Time Series Class-Incremental Learning (TSCIL) based on Pytorch. The paper has been accepted by SIGKDD 2024. Our CIL benchmarks are established with open-sourced real-world time series datasets. Based on these, our toolkit provides a simple way to customize the continual learning settings. Hyperparameter selection is based on Ray Tune.

What's new

  • Jun 2024: Include FastICARL into our toolkit.

  • May 2024: Our TSCIL paper has been accepted by SIGKDD 2024 (ADS track).

  • Feb 2024: Release of TSCIL toolkit.

Requirements

Create Conda Environment

  1. Create the environment from the file

    conda env create -f environment.yml

  2. Activate the environment tscl

    conda activate tscl

Dataset

Available Datasets

  1. UCI-HAR
  2. UWAVE
  3. Dailysports
  4. WISDM
  5. GrabMyo

Data Prepareation

We process each dataset individually by executing the corresponding .py files located in data directory. This process results in the formation of training and test np.array data, which are saved as .pkl files in data/saved. The samples are processed into the shape of (𝐿,𝐶).

For datasets comprising discrete sequences (UCI-HAR, Uwave and Dailysports), we directly use their original raw sequences as samples. For datasets comprising long-term, continuous signals (GrabMyo and WISDM), we apply sliding window techniques to segment these signals into appropriately shaped samples (downsampling may be applied before window sliding). If the original dataset is not pre-divided into training and testing sets, a manual train-test split will be conducted. Information about the processed data can be found in utils/setup_elements.py. The saved data are not preprocessed with normalization due to the continual learning setup. Instead, we add a non-trainable input normalization layer before the encoder to do the sample-wise normalization.

For convenience, we provide the processed data files for direct download. Please check the "Setup" part in the "Get Started" section.

Adding New Dataset

  1. Create a new python file in the data directory for the new dataset.
  2. Format the data into discrete samples in format of numpy array, ensuring each sample maintains the shape of (𝐿,𝐶). Use downsampling or sliding window if needed.
  3. If the dataset is not pre-divided into training and test subsets, perform the train-test split manually.
  4. Save the numpy arrays of training data, training labels, test data, and test labels into x_train.pkl, state_train.pkl,x_test.pkl, state_test.pkl in a new folder in data/saved.
  5. Finally, add the necessary information of the dataset in utils/setup_elements.py.

(back to top)

Continual Learning Algorithms

Existing Algorithms

Regularization-based:

Replay-based:

Adding New Algorithm

  1. Create a new python file in the agent directory for the new algorithm.
  2. Create a subclass that inherits from the BaseLearner class in agent/base.py.
  3. Customize methods including train_epoch(), after_task(), learn_task() and so on, based on your needs.
  4. Add the new algorithm to agents in agents/utils/name_match.py. If memory buffer is used, add it into agents_replay as well.
  5. Add the hyperparameters and their ranges for the new algorithm into config_cl within experiment/tune_config.py.

(back to top)

Getting Started

Setup

  1. Download the processed data from Google Drive. Put it into data/saved and unzip 
    cd data/saved
unzip <dataset>.zip
    You can also download the raw datasets and process the data with the corresponding python files.
  2. Revise the following configurations to suit your device:
    • resources in tune_hyper_params in experiment/tune_and_exp.py (See here for details)
    • GPU numbers in the .sh files in shell.

Run Experiment

There are two functions to run experiments. Set the arguments in the corresponding files or in the command line.

  1. Run CIL experiments with custom configurations in main.config.py. Note that this function cannot tune/change the hyperparameters for multiple runs. It is recommended for use in sanity checks or debugging.

    python main_config.py

  2. Tune the hyperparameters on the Val Tasks first, and then use the best hyperparameters to run experiment on the Exp Tasks:

    python main_tune.py --data DATA_NAME --agent AGENT_NAME --norm BN/LN

    To run multiple experiments, you can revise the script shell/tune_and_exp.sh and call it:

    nohup sh shell/tune_and_exp.sh &

    To reproduce the results in the paper, use the corresponding .sh files:

    nohup sh shell/{data}_all_exp.sh &

    We run the experiment for multiple runs to compute the average performance. In each run, we randomize the class order and tune the best hyperparameters. So the hyperparameters are different across runs. The searching grid of hyperparamteters is set in experiment/tune_config.py. Experiment results will be saved as log into result/tune_and_exp.

Custom Experiment Setup

Change the configurations in

  • utils/setup_elements.py: Parameters for data and task stream, including Number of tasks / Number of classes per task / Task split
  • experiment/tune_config.py: Parameters for main_tune.py experiments, such as Memory Budget / Classifier Type / Number of runs / Agent-specific parameters, etc.

For ablation study, revise the corresponding parameters in experiment/tune_config.py and rerun the experiments.

For online continual learning, set epochs to 1 and er_mode to online. (beta)

(back to top)

Acknowledgements

Our implementation uses the source code from the following repositories:

Contact

For any issues/questions regarding the repo, please contact the following.

Zhongzheng Qiao - qiao0020@e.ntu.edu.sg

School of Electrical and Electronic Engineering (EEE), Nanyang Technological University (NTU), Singapore.

(back to top)