This repo contains the code for the experiments in the paper: Weidong Geng, Yu Du, Wenguang Jin, Wentao Wei, Yu Hu, Jiajun Li. "Gesture recognition by instantaneous surface EMG images." Scientific Reports 6 (2016).

Please see http://zju-capg.org/myo for details.

• A CUDA compatible GPU
• Ubuntu 14.04 or any other Linux/Unix that can run Docker
• Docker
• Nvidia Docker

Following commands will (1) pull docker image (see docker/Dockerfile for details); (2) train ConvNets on the training sets of NinaPro DB1, CapgMyo DB-a and CSL-HDEMG, respectively; and (3) test trained ConvNets on the test sets.

Navigate to the downloaded version of this repo then run the following commands to use images from Docker Hub

mkdir .cache
# put NinaPro DB1 in .cache/ninapro-db1
# put CapgMyo DB-a in .cache/dba
# put CSL-HDEMG in .cache/csl
docker pull lif3line/sigr:latest
sudo scripts/trainsrep.sh
sudo scripts/testsrep.sh

Alternatively you can rebuild the images from the dockerfiles in this repo:

mkdir .cache
# put NinaPro DB1 in .cache/ninapro-db1
# put CapgMyo DB-a in .cache/dba
# put CSL-HDEMG in .cache/csl
cd docker/mxnet
sudo docker build .
# Note #ID of image on completion
cd ..

Edit the Dockerfile in this directory replacing FROM lif3line/mxnet:latest with FROM #ID:latest

sudo docker build .
# Note #ID of this new image
cd ..
cd scripts

Edit runsrep replacing lif3line/sigr:latest with #ID:latest

cd ..
sudo scripts/trainsrep.sh
sudo scripts/testsrep.sh

Training on NinaPro and CapgMyo will take 1 to 2 hours depending on your GPU. Training on CSL-HDEMG will take several days. You can accelerate training and testing by distribute different folds on different GPUs with the gpu parameter.

The NinaPro DB1 should be segmented according to the gesture labels and stored in Matlab format as follows. .cache/ninapro-db1/data/sss/ggg/sss_ggg_ttt.mat contains a field data (frames x channels) represents the trial ttt of gesture ggg of subject sss. Numbers are starting from zero. Gesture 0 is the rest posture. For example, .cache/ninapro-db1/data/000/001/000_001_000.mat is the 0th trial of 1st gesture of 0th subject, and .cache/ninapro-db1/data/002/003/002_003_004.mat is the 4th trial of 3th gesture of 2nd subject. You can download the prepared dataset from http://zju-capg.org/myo/data/ninapro-db1.zip or prepare it by yourself.

Licensed under an GPL v3.0 license.

Thanks DMLC team for their great MxNet!