EMG experiments

This is the code for the paper Signals to Spikes for Neuromorphic Regulated Reservoir Computing and EMG Hand Gesture Recognition, available here.

Getting started

Dependencies

numpy
pandas
matplotlib
scipy
scikit_learn
scikit_plot
nni
seaborn
Brian2

Installation

Requirements can be installed with :

pip3 install -r requirements.txt

Please note that the requirement file might use outdated dependencies. To get latest versions :

pip3 install numpy pandas matplotlib scipy scikit_learn scikit_plot nni seaborn Brian2

Finally, make sure the source of the project is in the Python path.

On Linux : export PYTHONPATH="$PWD/src"
On Windows : set PYTHONPATH="$PWD/src"

You can also modify it permanently : see here.

Usage

Datasets

Different datasets for Electromyography (EMG) are provided in the folder datasets/ :

Roshambo (classes: rock, paper, scissor)
5_class (classes: pinky, elle, yo, index, thumb)
Pinch (classes: Pinch1, Pinch2, Pinch3, Pinch4)

Pipelines

Spike encoder & Evaluation baseline

This is the first pipeline described in the paper. To execute it :

python3 src/evaluate_encoder.py [options]

Reservoir

This is the second pipeline described in the paper. To execute it :

python3 src/evaluate_reservoir.py [options]

Parameters

All the parameters are listed and detailed in the file src/args_emg.py. They can be specified as follow :

python3 <<script>> --dataset="5_class" --learning_algorithm="critical" --cbf=1

Scripting

Pipelines can also be used in Python scripts :

import random
import numpy as np
from evaluate_encoder import *
from evaluate_reservoir import *
from args_emg import args as my_args

# Get default arguments 
# Can be modified (eg. args.adaptiveProb = 1)
args = my_args()

# Fix the seed of all random number generator
seed = int(args.seed)
random.seed(seed)
np.random.seed(seed)

# Spike encoding & Evaluation baseline
svm_score_enc,firing_rate_enc,svm_score_baseline_enc = evaluate_encoder(args)

# Reservoir 
lda_score,lda_score_input,svm_linear_score,svm_linear_score_input,svm_score,svm_score_input,firing_rate,nbsynapses,nbneurons = evaluate_reservoir(args)

Reproduce the results

Experiments presented in the paper can be reproduced by executing the files in the folder experiments/. These files are for parameter exploration : a grid search algorithm is used to test different combinations of parameters. Files can be executed with :

python3 -m experiments.experiment_exploration_v10

Best results

Default parameters reproduce the best results presented in the paper. Note that a reservoir with 320 neurons is used for a 3 class problem, and 2048 neurons for a 5 class problem. Hence, for 5_class dataset, macrocolumn shape must be modified to [4,4,4].

Acknowledgements

University of Sherbrooke. NEuro COmputational & Intelligent Signal Processing Research Group (NECOTIS)

nikhil-garg/EMG_exp