Code for the paper "Modeling Electrical Motor Dynamics using Encoder-Decoder with Recurrent Skip Connection" (Accepted in AAAI 2020)

This is the github repository containing the code for the paper "Modeling Electrical Motor Dynamics using Encoder-Decoder with Recurrent Skip Connection" by Sagar Verma, Nicolas Henwood, Marc Castella, Francois Malrait, and Jean-Christophe Pesquet.

Project page

**Note: All the tables in the paper are for models trained for predicting individual quantities from three input quantities. It is possible to use the model for any input-ouput combinations by passing appropriate arguments. Also MAE is first computed on normalized output and then aggregated and normalized. Ignore sign in the reported MAE, RMSLE, and RMSE.

Requirements

The code has been tested on:

• 2xNvidia V100 GPU
• Ubuntu 18.04 LTS on 48 vCPUs and 186 GB of RAM
• Python 3.6.10
• Pytorch v1.4.0

Dataset

Motor Data

Run

Installation

git clone https://github.com/INRIA-OPIS/MotorNN.git
git checkout AAAI2020_release
pip install -r requirements.txt
pip install -e .

Download and extract dataset. Create weights and logs path.

To train a model use following

cd MotorNN
python motor_dynamics/summoner.py --gpu=0 --task=train --train_sim_dir={DATA_PATH}/train_sim/ --val_sim_dir={DATA_PATH}/val_sim/ --weights_dir={WEIGHTS_PATH} --logs_dir={LOGS_PATH} --model=deep_cnn --epochs=100 --batch_size=512 --lr=0.1 --inp_quants='voltage_d,voltage_q,speed' --out_quants='current_d' --stride=1 --window=100 --act=relu --loss=mse

Contact

For any queries, please contact

Sagar Verma: [email protected]