benchmarks

EGRU

Event based GRU was publised as a conference paper at ICLR 2023: Efficient recurrent architectures through activity sparsity and sparse back-propagation through time (notable-top-25%)

Code for the experiments and benchmarks presented in the paper are published in this directory.

Requirements

• Python 3.8.6
• Pytorch 1.21.1
  • pip3 install torch torchvision --extra-index-url https://download.pytorch.org/whl/cu113
• EvNN library
• Install all other requirements
  • pip install -r requirements.txt

Experiments

Change to benchmarks directory and add it to python path

cd benchmarks
export PYTHONPATH=.:$PYTHONPATH

Command for running sequential MNIST experiments

python smnist/smnist.py --train-epochs 200 --units 590 --batch-size 500 --rnn-type egru --use-output-trace --use-grad-clipping --grad-clip-norm 0.25  --cuda

Command for running the DVS Gesture experiments

python dvs_gesture/dvs128.py --data /tmp/dataset/ --cache /tmp/cache/ --logdir ./logs/ --batch-size 40 --units 795 --unit-size 1 --num-layers 1 --frame-size 128 --run-title egrud795_rerun --train-epochs 500 --frame-time 25 --rnn-type egru --learning-rate 0.0009975 --lr-gamma 0.8747 --lr-decay-epochs 56 --event-agg-method mean --use-cnn --use-all-timesteps --dropout 0.6321 --dropconnect 0.08134 --zoneout 0.2319 --pseudo-derivative-width 1.7 --threshold-mean -0.2465 --activity-regularization --activity-regularization-constant 0.01 --augment-data

Language modelling experiments

Scripts and instructions can be found in the lm directory. See LM Readme