This repository contains code related to the SuperGAN (supervised generative adversarial network) framework.
This project uses a Python virtual environment. In order to utilize this feature and avoid package headaches, please open the directory in a terminal.
Afterwards, run the following command: python3 -m venv supergan_env
Then, once inside of this directory, execute the following command: source supergan_env/bin/activate. This will make it
so that you are running Python in its own little virtual environment.
Next, when inside the virtual environment run the script using python setup.py, which will make
sure that everything is in order and install the relevant packages
Use source supergan_env/bin/activate and then run everything inside of that environment.
There are two configuration files SuperGAN reads.
The first, model.conf provides hyperparameters used for training. It also provides save locations for the generator and discriminator
in models.generator_filename and models.discriminator_filename. This file should be in .toml format. An example is provided in this
repository.
The second is a .toml file to be provided via a command line parameter. This toml file should have at minimum the following data:
data_file_path: The path to the dataset being trained on.classifier_path: The path to the pre-trained classifier.class_label: The class to generate.
As our network generates time-series data, the data must be in the form (num_samples, seg_length, num_channels). Since our network is supervised, the corresponding class labels are also required. Two versions must be included: 1) standard integer encoding 2) one-hot vector encoding. It is assumed that the data as well as standard and one-hot labels are included in a .h5 file and are named "X", "y" and "y_onehot" respectively.
Format is python main.py config_file.toml
The following command line parameters are accepted:
-h, --help: Display a help message-s, --save: Saves the Generator and Discriminator after training. This save location is provided inmodel.conf-S, --save_samples: Saves a number of generated samples of data after training.-l, --load: Loads a pre-trained GAN to generate samples from. The file location is provided inmodel.conf-C, --ignore_classifier: Trains the GAN without the classifier loss function.-R, --ignore_regularizer: Trains the gAN without the SFD regularization loss function.-c COUNT, --count COUNT: Specifies the number of samples to generate.
Our program requires the following packages: tensorflow, Keras, sklearn, numpy, matplotlib, h5py, toml
accuracy_epochs.rb : A script which is run after do_experiments.rb to obtain the average accuracy and number of training epochs for each
configuration of neither loss function, regularizer only, classifier only, and both loss functions.
CASAS_adlnormal_dataset.h5, sports_data_accelerometer.h5, and sports_data_gyroscope.h5 : Datasets created using the preprocessing scripts
in the Data Preprocessing repo.
compute_rtr_similarity.py : A script for calculating RTR similarity over some dataset and class label.
config_file_parser.py : Module for processing the model.conf file.
do_experiments.rb : A Ruby script which was used to automate experiments used in the paper.
Trains a model over every class and dataset.
example.toml, example_w_save.toml : Example .toml inputs to be provided via command line parameters.
extract_labels.rb : A script which is run after do_experiments.rb to obtain a CSV with all STS similarity scores presented
in our table of data.
gan_model.py : Module for constructing GAN model given configuration.
input_module.py : Contains necessary functions for processing the .toml input file and loading the appropriate data.
LSTM_accelerometer.h5, LSTM_adlnormal.h5, LSTM_gyroscope.h5 : Pre-trained classifiers for the three datasets.
main.py : Main file which takes the conditions from the .toml file and trains a generator for that given case.
model.conf : Configuration containing hyperparameters for training.
models.py: : Contains necessary models used in SuperGAN framework.
plotting_module.py : Contains necessary functions for displaying and saving plots of both real and generated data
saving_module.py : Contains necessary functions for saving training results and generator weights.
train_simple_lstm.py : Contains code for training the LSTM classifiers.
training_module.py : Functions for training generator and assessing data. In particular, contains functions for training
generator and discriminator, generating synthetic data, and computing the similarity metrics.
To produce the data presented in these tables, first run ruby do_experiments.rb, then, for each dataset, run
ruby extract_labels.rb dataset_name (dataset names are accelerometer, adlnormal, and gyroscope) to obtain
a CSV file containing all STS similarity values. The RTR similarity values can be obtained using the compute_rtr_similarity.py
script, which is run as follows: python3 compute_rtr_similarity.py -c class_label dataset_file.h5
STS Values for Accelerometer Dataset:
| 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | |
|---|---|---|---|---|---|---|---|---|---|
| 0.7612 | 0.7435 | 0.9997 | 0.9965 | 0.9993 | 0.9958 | 0.9995 | -0.0710 | 0.9959 | |
| R | 0.9424 | 0.9631 | 0.9389 | 0.9783 | 0.9562 | 0.9574 | 0.9715 | 0.9449 | 0.0173 |
| C | 0.2175 | 0.2736 | 0.0873 | 0.0241 | 0.2351 | 0.2930 | 0.0362 | 0.0334 | 0.2286 |
| CR | 0.1572 | 0.2046 | 0.1506 | 0.0454 | 0.0908 | 0.1941 | 0.1665 | 0.0960 | 0.0610 |
| RTR | 0.0063 | 0.0075 | 0.0022 | 0.0014 | 0.0031 | 0.0016 | 0.0015 | 0.0019 | 0.0021 |
STS Values for Gyroscope Dataset:
| 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | |
|---|---|---|---|---|---|---|---|---|---|
| 0.9935 | 0.9950 | 0.8631 | 0.9985 | 0.8625 | 0.9974 | 0.7108 | 0.9961 | 0.9960 | |
| R | 0.9524 | 0.9679 | 0.9325 | 0.9955 | 0.9940 | 0.0161 | 0.9716 | 0.9931 | 0.8438 |
| C | 0.0937 | 0.0496 | 0.0250 | 0.3785 | 0.1878 | 0.0327 | 0.2404 | 0.1271 | 0.0302 |
| CR | 0.0267 | 0.4181 | 0.1065 | 0.4279 | 0.1306 | 0.1195 | 0.1001 | 0.2564 | -0.0195 |
| RTR | 0.6610 | 0.7779 | 0.3019 | 0.8334 | 0.8262 | 0.6446 | 0.7801 | 0.8865 | 0.3569 |
STS Values for CASAS ADL Dataset:
| 0 | 1 | 2 | 3 | 4 | |
|---|---|---|---|---|---|
| 0.9931 | 0.9819 | 0.9740 | 0.9692 | 0.9830 | |
| R | 0.1048 | 0.0551 | 0.5977 | 0.0014 | 0.1567 |
| C | 0.0898 | 0.1060 | 0.1820 | 0.0698 | 0.2372 |
| CR | 0.0965 | 0.1391 | 0.1382 | 0.0273 | -0.0671 |
| RTR | 0.0032 | 0.0164 | 0.1518 | 0.0023 | 0.0382 |