KalmanHD: Robust On-Device Time Series Forecasting with Hyperdimensional Computing

This repo contains the simulation implementation for paper:

Ivannia Gomez, Xiaofan Yu, Tajana Rosing. "KalmanHD: Robust On-Device Time Series Forecasting with Hyperdimensional Computing" in the Proceedings of ASP-DAC 2024.

Paper PDF

File Structure

.
├── Stuff              // DatasetLoader and libraries necesarry for running model
├── images             // Different image results of the models
├── models             // Implementation of ML models - Baselineas + KalmanHD
├── preprocessed_data  // Datasets data
├── scripts            // Collection of bash and .py scripts for various experiments in the paper
├── README.md          // This file
├── main.py            // Main script that runs and tests each model
├── plot.ipynb         // Notebook to plot the accuracy results of the different models in the 3 different types of noise
├── plot_eff.ipynb     // Notebook to plot the running times results of the different models
└── requirements.txt   // Prerequisites

Prerequisites

We test with Python3.9. We recommend using conda environments:

Steps to setup the repository:

1. Clone this repository.

   git clone https://github.com/DarthIV02/KalmanHD

2. Create a new anaconda environment and activate.

   conda create -n hd_forecasting python=3.9
   conda activate hd_forecasting

3. Run the following commands to install python packages:

   pip install -r requirements.txt
   pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu117

All Python packages required are included in requirements.txt and can be installed automatically except for torch with CUDA (in the case of running script in GPU).

Dataset Preparation

As mentioned in the paper, we experiment on Energy Consumption Fraunhoufer, San Francisco Traffic, Metro Interstate Traffic Volume, Guangzhao Traffic and Electricity Load Diagrams from Tzagkarakis, C. et. al.

• All the dataset data files can be found in ./preprocessed_data folder (most of them are .csv files).
• The datasets are loaded into the main script with the help of ./Stuff/DatasetLoader.py file.

Getting Started

To run KalmanHD without any type of noise in the dataset SanFranciscoTraffic:

python main.py --model KalmanHD --dataset SanFranciscoTraffic

To run any other baseline, run the past command with the attribute "--model" changed:

• RegHD: Implementation of RegHD in RegHD: Robust and Efficient Regression in Hyper-Dimensional Learning System, DAC '21
• VAE: Implementation of PFVAE in PFVAE: a planar flow-based variational auto-encoder prediction model for time series data, Mathematics 2022
• DNN: Implementation of E-Sense in A diverse noise-resilient dnn ensemble model on edge devices for time-series data, SECON '21
• KalmanFilter: Implementation of online Kalman Filter in Autoregressive-model-based methods for online time series prediction with missing values: an experimental evaluation

To run any another dataset, run the past command with the attribute "--dataset" changed:

• SanFranciscoTraffic: San Francisco Traffic [default if no dataset specified]
• MetroInterstateTrafficVolume: Metro Interstate Traffic Volume
• GuangzhouTraffic: Guangzhao Traffic
• EnergyConsumptionFraunhofer: Energy Consumption Fraunhoufer
• ElectricityLoadDiagrams: Electricity Load Diagrams

To change the learning rate in the HD-based methods change the "--learning_rate" attribute with the float-value desired as shown below (default: 0.0001):

python main.py --model KalmanHD --learning_rate 0.0001

We list other hyperparameters that can be modified:

• --dimension_hd: To change the dimension of the hypervectors in the HD-based methods with the int-value desired (default: 1000)
• --hd_encoder: To change the encoder function in the HD-based methods (default: 'nonlinear') other options include:
  • 'linear': Linear-based encoding
  • 'time_encoding': TimeSeries-based encoding, with permutation on each time stamp
• --print_freq: How often to test the current model with the cross-validation data (default: 1000)
• --epochs: Number of training epochs or number of passes on dataset (default: 1). All of the experiments are set to 1, given that we test on the online setting.
• --levels: Number of levels to divide encoding when using a linear hd-encoder for HD-based methods (default: 6)
• --size_of_sample: Number of previous samples before forecasting (default: 20)
• --alpha: Percentage of importance of the past variance in the moving average variance (default: 0.3)
• --device: Device to use if possible, either cpu or gpu (default: 'gpu')

We list the arguments that can be modified for the different types of noises:

• Gaussian Noise:
  • --gaussian_noise: Standard deviation of the gaussian noise (default: 0.0). Change this parameter to any other float-value to add gaussian noise.
• Poisson Noise:
  • --poisson_noise: Lambda for the poisson noise (default: 0.0). Change this parameter to any other float-value to add poisson noise.
• Missing Values:
  • --s: Number of consecutive missing values (default: 5)
  • --p: Percentage of probability of value missing (default: 0.0). Change this parameter to any other float-value between (0-1) to add missing values. The higher the porcentage, the higher number of missing values in the input training dataset.

Scripts

We provide our scripts for running various experiments in the paper in scripts:

.
├── run_exp.py                           // Python script for running baselines in all of the datasets with the hyperparameters set in the paper
├── run_exp.sh                           // Bash script for running baselines in all of the datasets with the hyperparameters set in the paper
├── run_exp_gaussian.py                  // Python script for running all of the methods with Gaussian noise added
├── run_exp_gaussian.sh                  // Bash script for running all of the methods with Gaussian noise added
├── run_exp_missing.py                   // Python script for running all of the methods with different levels of missing data
├── run_exp_missing.sh                   // Bash script for running all of the methods with different levels of missing data
├── run_exp_poisson.sh                   // Bash script for running all of the methods with Possion noise added
└── run_test_lr.sh                       // Bash script for testing different levels of learning rate in KalmanHD on all datasets