Cooperative Neural Networks(CoNN) : Exploiting prior independence structure for improved classification
(also known as Joint Constraint Networks).
This work is published in NIPS 2018 link.
I am working on a Blog to highlight the novelties and main contributions of this work.
The file 'main.py' is tested on the following
- Python 3
- Pytorch 0.2.0
- Numpy 1.15.1
- nltk (preprocessing text)
- P100 GPUs
Getting the data link:
$ wget https://www.cs.jhu.edu/~mdredze/datasets/sentiment/unprocessed.tar.gz
$ tar -xvzf unprocessed.tar.gz
Running the script:
$ python main.py
with the default argparse settings, I get roughly
Avg auc(5 fold CV) = 0.9213456403902894 with std dev = 0.007758676219792392
I will be updating the repo with the following additions
- Script compatible with latest Pytorch version
- Script for latest Tensorflow version
Issues can be reported at issues section.
If you want to discuss or contribute, please feel free to drop a mail or raise an issue :)
I will be happy to discuss and collaborate, if you want to use CoNN or its variant for some other Graphical models!
CoNN is released under Apache License. You can read about our license at here