This repository is an artifact for the paper under review "Multivariate Generative Adversarial Networks and their Loss Functions for Synthesis of Multichannel ECGs" submitted to IEEE Pattern Recognition and Machine Intelligence 2020. The code is available as Jupyter Notebooks .ipynb, .py files will be added shortly.
Run through the Notebooks Arr_Data_Prep.ipynb and Nsr_Data_Prep.ipynb to download and preprocess the datasets
See above
train.py
Trains the GAN model defined in model.py and saves each generator and discriminator after every training epoch as .pt files.
The evaluation metrics are comptued for a batch of data from one of the generators of your choice. You can choose to generate a batch of data and use it for evaluation.
mmd_eval.py
Computes MMD 3 Sample Test from mmd.py (taken from eugenium's Github), returns a test statistic for whether the training data is closer to the generated data then a random noise sample.
dtw_eval.py
Calculates the dependent multivariate Dynamic Time Warping values between the training and generated data.
membership_inference_attack.py
Computes a simple membership inference attack on the generated data from the model. It uses the euclidean distance for the multivariate signals. You can use the any of the DTW functions as the distance in mia_dtw.py but the time taken to execute varies significantly.
The training losses and generated results are plotted using the Notebook plot_gen.ipynb