TactileLocNeurons

This package is a PyTorch implementation of the paper "Event-Driven Tactile Learning with Location Spiking Neurons".

TSRM vs. LSRM

The hybrid model

Citation

Kang, Peng and Banerjee, Srutarshi and Chopp, Henry and Katsaggelos, Aggelos and Cossairt, Oliver. "Event-Driven Tactile Learning with Location Spiking Neurons". In 2022 International Joint Conference on Neural Networks (IJCNN 2022).

@inproceedings{kangTactile,
        title={Event-Driven Tactile Learning with Location Spiking Neurons},
        author={Kang, Peng and Banerjee, Srutarshi and Chopp, Henry and Katsaggelos, Aggelos and Cossairt, Oliver},
        booktitle={2022 International Joint Conference on Neural Networks (IJCNN)},
        pages={1--8},
        year={2022},
        organization={IEEE}
}

Requirements

Python 3 with slayerPytorch and the packages in the requirements.txt:

The project has been tested with one RTX 3090 on Ubuntu 20.04 / Ubuntu 22.04. The training and testing time should be in minutes.

Installation

1. Follow the requirements and installation of slayerPytorch to install it, see slayerPytorch/README.md.
2. Install any necessary packages in the requirements.txt with pip install or conda install.

Datasets

1. Donwload the preprocessed data here.
2. Save the preprocessed data for Objects, Containers, and Slip Detection in datasets/preprocessed.

Training

python locsnn/train_location_snn.py --epoch 500 --lr 0.001 --sample_file 1 --batch_size 8 --fingers both --data_dir <preporcessed data dir> --hidden_size 32 --loss NumSpikes --mode location --network_config <network_config>/container_weight_location.yml  --task cw --checkpoint_dir <checkpoint dir>

Experiments

1. The hybrid model with the whorl-like location order:

python locsnn/train_location_snn.py --epoch 500 --lr 0.001 --sample_file 1 --batch_size 8 --fingers both --data_dir <preporcessed data dir> --hidden_size 32 --loss NumSpikes --mode location_cat_whorl --network_config <network_config>/container_weight_location.yml  --task cw --checkpoint_dir <checkpoint dir>

2. Location Tactile SNN:

python locsnn/train_location_snn.py --epoch 500 --lr 0.001 --sample_file 1 --batch_size 8 --fingers both --data_dir <preporcessed data dir> --hidden_size 32 --loss NumSpikes --mode only_location --network_config <network_config>/container_weight_location_only.yml --task cw --checkpoint_dir <checkpoint dir>

3. $\lambda$ tuning in the weighted loss function: $\lambda$ value can be changed in slayerPytorch/src/spikeLoss.py, but remember to install slayerPytorch again to activate the changes.

python locsnn/train_location_snn.py --epoch 500 --lr 0.001 --sample_file 1 --batch_size 8 --fingers both --data_dir <preporcessed data dir> --hidden_size 32 --loss WeightedLocationNumSpikes --mode location --network_config <network_config>/container_weight_location.yml --task cw --checkpoint_dir <checkpoint dir>

4. Confusion matrices on Containers

python confusion/confusion_location.py --runs <checkpoint dir>/cw_location_1

5. Timestep-wise inference

python timestep_inference/inference_timestep.py --runs <checkpoint dir>/cw_location_1 --save <timestep inference dir>

Trained model examples

1. Download models from [https://drive.google.com/drive/folders/1XBzpbk5Vt7E7qevlOW06GvFY0N_N8ymU?usp=sharing].
2. Save the models in history folder.

Troubleshooting

if your scripts cannot find the locsnn module, please run in the root directory:

export PYTHONPATH=.

Credits

The codes of this work are based on slayerPytorch and VT-SNN.