GrOWL

Learning to share: simultaneous parameter tying and sparsification in deep learning

Dejiao Zhang*, Haozhu Wang*, Mario Figueiredo, Laura Balzano (*Co-first author)
https://openreview.net/pdf?id=rypT3fb0b

To run the code:

1. Compile the c code in ./owl_projection by the following:

gcc -fPIC -shared -o libprox.so proxSortedL1.c

2. VGG on Cifar10 (VGG folder)

python vgg_main.py 

You can switch to different regularizers by changing the configuration info in "flags.py"  
All configuration settings are in vgg_main.py and flag.py  
Some of the hyperparameters are described in the table in our paper, e.g., preference value

3. Fully connected network on MNIST (MNIST folder)

python run_exp.py (reproduce results in table 1)

All available configuration settings are contained in experiment_config folders,   
you can modify the settings either in the .yaml file "10_20_config_search.yaml" or in the "run_exp.py" script.

4. Plots (Plots folder)

python gen_fig4.py (generate figure 4)
python gen_fig5.py (generate figure 5)
python gen_fig6.py (generate figure 6)

To run the the above codes successfully, please modify the "file_names" and "log_root"   
in the code to match your local files. 

Dependencies:

Tensorflow 1.0.0
Numpy 1.14.0
Scipy 1.0.0
Matplotlib 2.1.0
Scikit-learn 0.19.1

Our experiments were done on Ubuntu 16.04.3 LTS.