DJINN: Deep jointly-informed neural networks
DJINN is an easy-to-use algorithm for training deep neural networks on supervised regression tasks. For additional information, refer to the paper "Deep neural network initialization with decision trees", cited below.
DJINN requires TensorFlow (v1.0.1 or later) and
scikit-learn (v0.18 or later is recommended).
DJINN also uses numpy, matplotlib, and cPickle.
Sphinx is required to view the html documentation.
Note that the sklearn version used when training a DJINN model must be the same version used when reloading/evaluating the saved model.
To use DJINN, clone the repo and install:
$ git clone https://github.com/LLNL/DJINN.git
$ cd DJINN
$ pip install -r requirements.txt
$ pip install .
Try it out! Examples for training DJINN models are included in the tests folder.
-python djinn_example.py (single output)
-python djinn_multiout_example.py (multiple outputs)
For Mac users with Anaconda installs, it might be necessary to manually install matplotlib via pip:
$ pip install matplotlib
If matplotlib will not import, try running "pythonw", for example:
$ pythonw djinn_example.py
To view the DJINN documentation:
cd docs
make html
Open docs/_build/html/index.html in a browser
DJINN is available at https://github.com/LLNL/DJINN
If you use DJINN in your research, please cite the following paper:
"Deep neural network initialization with decision trees", Humbird, K. D.; Peterson, J. L.; McClarren, R. G., IEEE Transactions on Neural Networks and Learning Systems (2018). Early access: 10.1109/TNNLS.2018.2869694
Copyright (c) 2018, Lawrence Livermore National Security, LLC.
Produced at the Lawrence Livermore National Laboratory
Written by K. Humbird ([email protected]), L. Peterson ([email protected]).
LLNL-CODE-754815 OCEC-18-117
All rights reserved.
Unlimited Open Source- BSD Distribution.
For release details and restrictions, please read the RELEASE, LICENSE, and NOTICE files, linked below: