LifeHD

[IPSN 2024] Lifelong Intelligence Beyond the Edge using Hyperdimensional Computing

File Structure

.
├── LICENSE
├── README.md           // this file
├── main.py             // main file
├── methods             // implementation of LifeHD
├── requirements.txt
├── scripts             // scripts to run each method
└── utils               // utils files

Prerequisites

We test with Python3.8. We recommend using conda environments:

conda create --name lifehd-py38 python=3.8
conda activate lifehd-py38
python3 -m pip install -r requirements.txt

All require Python packages are included in requirements.txt and can be installed automatically.

Dataset Preparation

As mentioned in the paper, we experiment on MHEALTH, ESC-50 and CIFAR-100. This repo also includes the dataset setup for MNIST, CIFAR-10, HAR (HAR-timeseries) and ISOLET.

We preprocess the MHEALTH and HAR-timeseries data from their original data. HAR-timeseries is the raw time series version of the dataset. To use, please download the corresponding files and place them in the appropriate path.
- To use MHEALTH
```
mkdir datasets # create the datasets folder if it does not exist
mkdir datasets/MHEALTH # create the folder for MHEALTH
```
  Download the preprocessed MHEALTH data, mhealth.mat, and place it inside the MHEALTH folder.
- To use HAR-timeseries
```
mkdir datasets # create the datasets folder if it does not exist
mkdir datasets/HAR_TimeSeries # create the folder for HAR-timeseries
```
  Download the preprocessed HAR time series data, har.mat, and place it inside the HAR_TimeSeries folder.
We adapt the ESC-50 data. To simplify the setup, we offer our processed version. Please download esc50.zip and place it under the following newly created folder
```
mkdir datasets/esc50 # create the folder for ESC-50
```

The rest datasets should be downloaded and configured automatically by our group.

Model Preparation

In the HDnn encoding, we need a pretrained and frozen neural network as feature extractor. For CIFAR-10/CIFAR-100, we offer pretrained ResNets and MobileNets downloaded from torchvision. For ESC-50, we employ the pretrained ACDNet adapt from their Github repo.

To setup these pretrained model, please download pretrained_models.zip and unzip it under utils/.

Getting Started

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

To run the supervised HDC
- MNIST/CIFAR-10/CIFAR-100/ESC-50/HAR
```
bash run_basichd.sh BasicHD <cifar10/cifar100/mhealth/esc-50/har> <iid/seq> <trial ID> idlevel
```
<iid/seq> specifies the data stream order. idlevel configures the encoding method, which indicates a simpler version of the spatiotemporal encoding.

For example, trial 0 of the CIFAR-10, class-incremental streams experiment could be fired with
```
bash run_basichd.sh BasicHD cifar10 seq 0 idlevel
```
- MHEALTH/HAR_timeseries
```
bash run_basichd.sh BasicHD <mhealth/har_timeseries> <iid/seq> <trial ID> spatiotemporal
```
The difference is the last argument - we set to use the full spatiotemporal encoding here.

To run LifeHD

MNIST/CIFAR-10/CIFAR-100/ESC-50

bash run_lifehd.sh LifeHD <cifar10/cifar100/mhealth/esc-50> <iid/seq> <trial ID> idlevel

MHEALTH/HAR_timeseries

bash run_lifehd_timeseries.sh LifeHD <mhealth/har_timeseries> <iid/seq> <trial ID> spatiotemporal

To run SemiHD

MNIST/CIFAR-10/CIFAR-100/ESC-50

bash run_semihd.sh SemiHD <cifar10/cifar100/mhealth/esc-50> <iid/seq> <trial ID> idlevel <label_ratio>

MHEALTH/HAR_timeseries

bash run_semihd.sh SemiHD <mhealth/har_timeseries> <iid/seq> <trial ID> spatiotemporal <label_ratio>

The commands are very similar to the previous cases, except adding an additional argument for labeling ratio in the end.

To run LifeHDsemi

MNIST/CIFAR-10/CIFAR-100/ESC-50

bash run_lifehdsemi.sh LifeHDsemi <cifar10/cifar100/mhealth/esc-50> <iid/seq> <trial ID> idlevel <label_ratio>

MHEALTH/HAR_timeseries

bash run_lifehdsemi.sh LifeHDsemi <mhealth/har_timeseries> <iid/seq> <trial ID> spatiotemporal <label_ratio>

To run LifeHDa LifeHDa can be fired with the same scripts mentioned above by appropriately configuring the folloiwng arguments in each script:
- --mask_mode: fixed or adaptive
- --mask_dim: dimension of the mask

License

MIT

If you have any questions, please feel free to contact x1yu@ucsd.edu.