Code of AccelGSCAD

This is the Cython code of Group SCAD/Fast Group SCAD/AccelGSCAD.

Files

• README.md
• setup.py
• skip_setup.py
• fast_setup.py
• group_scad.pyx
• skip_group_scad.pyx
• fast_group_scad.pyx
• sgl_tools.py
• experiment.py
• grid_experiments.sh
• license.txt

Environment

The environment can be made by using the Dockerfile of Kaggle as follows:

https://github.com/Kaggle/docker-python

You may need to write the following additional lines in the Dockerfile.

RUN pip install ipdb
RUN conda install -c anaconda cython
RUN pip3 install numpy
RUN pip3 install scipy
RUN pip3 install sklearn

We recommend you to use Python 3.8 and Cython 2.9.

Datasets

• eunite2001
  • You can download the dataset from LIBSVM cite.
  • Please put the dataset under ./data.

https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/regression/eunite2001

• triazines
  • You can download the dataset from LIBSVM cite.
  • Please put the dataset under ./data.

https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/regression/triazines_scale

• The other datasets are automatically downloaded from OpenML cite.

Compile

We can compile our cython codes on the above environment as follows:

• Group SCAD (Breheny et al., Statistics and Computing, 2015.)

python3 setup.py build_ext --inplace

• Fast Group SCAD (Ida et al., NeurIPS, 2019.)

python3 skip_setup.py build_ext --inplace

• AccelGSCAD (ours)

python3 fast_setup.py build_ext --inplace

Usage

• Please perform grid_experiments.sh on the above docker environment.

bash grid_experiments.sh

• The above command generates files of the processing times, the objective values, the losses, the parameters and the logs.

• Each method on the datasets of eunite and qsbralks can be performed in less than a minute on one CPU core of 2.20 GHz Intel Xeon server running Linux.

• Each method on the datasets of qsbr_rw1, qsf and triazines is commented out in the code of grid_experiments.sh because it takes more than 10 minuetes. Please see the code if you want to perform the methods on these datasets.

Note

• If you want to change the hyperparameter m in our method, you can rewrite line 129 in the code of fast_group_scad.pyx and re-compile it.
• If you want to evaluate the exact processing time, please specify the core ID of the CPU (logical core, not physical core) on docker, e.g. --cpuset-cpus=0,20. In this case, please specify the logical core IDs so that they do not cross the physical cores.