creg

Fast regression modeling framework.

Building

To build creg, you will need:

• Download the 2014-02-18 version of the source code (or install from github).
• If you are working with the code from github
  • Run autoreconf -fvi
• If you are working with a source tarball
  • Run tar xzf creg-2014-02-18.tar.gz
  • Change into the directory cd creg-2014-02-18
• Last, run the ./configure command and then type make.

Software requirements:

• A relatively recent C++ compiler (g++ or clang)
• The Boost C++ libraries installed somewhere

Examples

Logistic regression example (training only):

./creg/creg -x test_data/iris.trainfeat -y test_data/iris.trainresp --l1 1.0 > weights.txt

• To load initial values for weights from a file (warm start), use -w FILENAME.

Logistic regression example (training and testing):

./creg/creg -x test_data/iris.trainfeat -y test_data/iris.trainresp --l1 1.0 \
     --tx test_data/iris.testfeat --ty test_data/iris.testresp > weights.txt

Logistic regression example (training and prediction):

./creg/creg -x test_data/iris.trainfeat -y test_data/iris.trainresp --l1 1.0 --tx test_data/iris.testfeat

• By default, the test set predictions and learned weights are written to stdout.
• If -D is specified, the full posterior distribution over predicted labels will be written.
• To write weights to a file instead of stdout, specify --z FILENAME. To suppress outputting of weights altogether, supply the -W flag.

Linear regression example (training and testing):

./creg/creg -n -x test_data/auto-mpg.trainfeat -y test_data/auto-mpg.trainresp --l2 1000 \
     --tx test_data/auto-mpg.testfeat --ty test_data/auto-mpg.testresp > weights.txt

Ordinal regression example (training and testing)

./creg/creg -o -x test_data/shuttle.trainfeat -y test_data/shuttle.trainresp \
    --tx test_data/shuttle.testfeat --ty test_data/shuttle.testresp > weights.txt

Note: for ordinal regression, labels have to be consecutive and start from 0 (e.g., 0/1/2 for 3 labels).

Data format

Training and evaluation data are expected to be in the following format:

• A feature file containing lines of the form:

    id1\t{"feature1": 1.0, "feature2": -10}
    id2\t{"feature2": 10,  "feature3": 2.53}
  

  where the JSON-like map defines a sparse feature vector for each instance

• A response file containing the same number of lines of the form:

    id1\t10.1
    id2\t4.3
  

  where the response is numeric for linear and ordinal regression and a label for logistic regression

You will find example files for each type of model in the test_data directory.

Python module

Quick install:

pip install -e git+https://github.com/redpony/creg.git#egg=creg

Some documentation is available on the wiki.