Our RegCGAN is a new generator for producing virtual samples.
Inspiring by CGAN, RegCGAN implicitly capture the condition
probability p(y|x) from historical data as like CGAN doing for
p(x|y).So, similar to CGAN, our RegCGAN is mainly made up of
a Generator G and a Discriminator D. The Generator G
consumes x and z and yields fake y, while
the Discriminator D consumes x and y and distinguish between
the true y and the fake one. Once RegCGAN is well trained,
it can serve as a generative probability model p(y|x), just like
Gaussian process (GP). Thanks to such ability, we make an attempt to
use it to synthesis output of any targeting points in the input space
For small sample size (SSS) soft modeling problems, data at hand suffers
from data sparsity, leading to degrading performance of a soft model.
To handle this issue, we intend to create new points to fill up such
areas of data sparsity in the input space through CVT sampling. The
data sparsity regions is identified by Local Outlier Factor. The
output of uniformly distributed new samples is synthesis by averaging
a number of samples drawn from p(y|x), which is captured by RegCGAN.
Because the generated samples has a similar behavior to the
real samples when used to training a soft model,
we call them virtual samples.
This code requires the following package referenced in requirements.txt:
- numpy, jupyter, matplotlib, pandas, seaborn
- sklearn
- tensorflow, keras
- density, available at
https://pypi.org/project/diversipy/ - idaes-pse, available at
https://idaes-pse.readthedocs.io/en/stable/getting_started.html/ - For GPy, one can visit
https://sheffieldml.github.io/GPy/
all packages need to be installed on a conda environment with python >= 3.0
- First, install must-have packages in the environment
pip install -r requirements.txt. - Second, install RegCGAN itself by
typing
pip install -e .at the root. - Finally, run an applications in the notebooks using
jupyter-notebook.
We appreciate efforts in https://github.com/eriklindernoren/Keras-GAN and
in https://github.com/mkirchmeyer/ganRegression.
Please cite in the way following:
Zhongsheng Chen, Kunrui Hou, Meiyu Zhu, Yuan Xu, Qunxiong Zhu.
A virtual sample generation approach based on
a modified conditional GAN and centroidal Voronoi tessellation sampling
to cope with small sample size problems:
application to soft sensing for chemical process. Applied Soft Computing,
Volume 101, March 2021, 107070. DOI:10.1016/j.asoc.2020.107070.