This is the PyTorch implementation of the paper: Hebbian Learning based Orthogonal Projection for Continual Learning of Spiking Neural Networks (ICLR 2024). [openreview] [arxiv]
- Python 3 (Recommend to use Anaconda)
- PyTorch, torchvision
- NVIDIA GPU + CUDA
- Python packages:
pip install numpy opencv-python
Run as following examples:
python spiking_train_pmnist.py -data_dir path_to_data_dir -out_dir log_checkpoint_name -gpu-id 0
# feedback alignment
python spiking_train_pmnist.py -data_dir path_to_data_dir -out_dir log_checkpoint_name -feedback_alignment -gpu-id 0
# sign symmetric
python spiking_train_cifar.py -data_dir path_to_data_dir -out_dir log_checkpoint_name -sign_symmetric -gpu-id 0
# baseline, i.e. vanilla sequential learning of different tasks
python spiking_train_fivedataset.py -data_dir path_to_data_dir -out_dir log_checkpoint_name -baseline -gpu-id 0
# combination with memory replay (if combined with -baseline, corresponds to only memory replay)
python spiking_train_fivedataset.py -data_dir path_to_data_dir -out_dir log_checkpoint_name -replay -gpu-id 0
# hlop with lateral spiking neurons
python spiking_train_pmnist.py -data_dir path_to_data_dir -out_dir log_checkpoint_name -hlop_spiking -hlop_spiking_scale 20. -hlop_spiking_timesteps 40 -gpu-id 0
# for convolutional networks, can specify the hlop projection type for acceleration on CPU/GPU
-hlop_proj_type weight
The default hyperparameters are the same as the paper.
Some codes are adpated from DSR, OTTT, and spikingjelly. Some codes for data processing are adapted from GPM.
If you have any questions, please contact [email protected].