Skip to content

AILAB-CEFET-RJ/HBRKGA

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

13 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

HBRKGA

HBRKGA is a hyperparameter optimization method for Neural Networks based on BRKGA framework (http://mauricio.resende.info/src/brkgaAPI/). We have done experiments comparing HBRKGA with some hyperparameter optimization methods. The results and description of our proposal can be found at https://arxiv.org/abs/2011.11062.

hbrkga_flux

Requirements

Python 3.6 Tensorflow 1.13.1 Numpy 1.16.13 Pandas 0.24.2 SciKit Learn 0.21.1 Matplotlib 3.2.2 C++ STL

Datasets

All datasets are publicly available at https://doi.org/10.5281/zenodo.4252922. Datasets must be placed in the data folder.

Usage

Install/load the requirements packages.

To setup the dataset, go to the C++ to python bridge in evaluate function at data and specifies one dataset:

python ./src/python/neural_network.py fashion (Fashion MNIST)
python ./src/python/neural_network.py mnist (MNIST)
python ./src/python/neural_network.py rectangles (Rectangles)
python ./src/python/neural_network.py cosmos (COSMOS)

This process can be also add to python data file with another trainning/validation set.

To setup the hyperparameters go to data and edit/add the Sample Decoder par_dom. Each par_dom represents a hyperparameter with min/max value to optimize.

	par_dom.push_back(make_pair(1000,2000));
	par_dom.push_back(make_pair(2000,4000));
	par_dom.push_back(make_pair(2000,6000));
	par_dom.push_back(make_pair(0.000001,0.1));
	par_dom.push_back(make_pair(0,0.001));

To compile the C++ project:

make clean
make

To execute the HBRKGA optimization process:

nohup ./bin/hypAG > hbrkga.log &

Citation

@article{serqueira2020population,
  title={A Population-based Hybrid Approach to Hyperparameter Optimization for Neural Networks},
  author={Serqueira, Marcello and Gonz{\'a}lez, Pedro and Bezerra, Eduardo},
  journal={arXiv preprint arXiv:2011.11062},
  year={2020}
}

Contact

To give your opinion about this work, send an email to [email protected].