snn-localization repo

PyTorch implementation of Spiking Neural Networks trained on surrogate gradient & BPTT using snntorch.

Install Dependencies

Originally runs on Python 3.6.13. Example of requirement.txt given. Includes:

conda install pytorch torchvision cudatoolkit=10.2 -c pytorch
pip install albumentations==0.4.1 snntorch tensorboard

Usage

Run python train_localization_SG.py -h to see all options available.

Here's an example:

python train_localization_SG.py "data/oxford_iiit_pet" --experiment="Name_Of_Experiment" --height=176 --width=240 --timesteps=40 --epochs=100 --batch-size=4 --print-freq=100 --neural-coding=rate --lr=1e-3 --save-sample=42

A folder will be created at location experiments/Name_Of_Experiments with all data related to the experiment. It can include:

1. Tensorboard logs in experiments/Name_Of_Experiments/logs (val & train IoU metric and losses ).
2. Prediction samples from the test set at each epoch in experiments/Name_Of_Experiments/samples
3. Videos of SAM activations in experiments/Name_Of_Experiment/SAMs

In addition, a testing + additional metrics can be obtained (SAMs) by running the following command:

python test_localization_SG.py "data/oxford_iiit_pet" --experiment="Name_Of_Experiment" --height=176 --width=240 --timesteps=40 --epochs=100 --batch-size=4 --print-freq=100 --neural-coding=rate --lr=1e-3 --save-sample=42