The HDC utilities is implemented using Cython. First run python setup.py build_ext --inplace to compile the cython code.
Download the MNIST, CIFAR10, CIFAR100 data from here. Place the data in HD-Glue/data.
We first train the end-to-end classifier and use the classifier as the image encoder. For example: python train_enc.py -net vgg11 -dataset CIFAR10 -gpu. This experiment only supports CIFAR10 and CIFAR100. The network can be chosen from vgg11, vgg13, vgg16, vgg19, resnet18, resnet34. To all possible combinations, please refer to train_enc.sh.
The simplest way to train is sh train_enc.sh. If you don't have a gpu, you should remove the -gpu argument.
If you don't want to train the network by yourself, you could download the pretrained models here and place them to HD-Glue/networks.
After obtaining the neural networks, run python gen_emb.py -dataset CIFAR10 and python gen_emb.py -dataset CIFAR100 to obtain the embeddings given by each model. The embeddings will be saved to HD-Glue/emb/.
If you don't want to compute the embeddings by yourself, you can download the pre-computed embeddings here and place them to HD-Glue/emb/.
After saving the embeddings, run e.g. python train_HD.py -size 1000 -dim 8000 -dataset CIFAR10 to get the experiment results. To see all the combinations, you can check and run sh train_HD.sh.
The HDC utilities with GPU support is under development. If you use our project, please cite:
@inproceedings{sutor2022gluing,
title={Gluing neural networks symbolically through hyperdimensional computing},
author={Sutor, Peter and Yuan, Dehao and Summers-Stay, Douglas and Fermuller, Cornelia and Aloimonos, Yiannis},
booktitle={2022 International Joint Conference on Neural Networks (IJCNN)},
pages={1--10},
year={2022},
organization={IEEE}
}