Jensen: A toolkit with API support for Convex Optimization and Machine Learning For further documentation, please see https://arxiv.org/abs/1807.06574

Copyright (C) Rishabh Iyer, John T. Halloran, and Kai Wei Licensed under the Open Software License version 3.0 See COPYING or http://opensource.org/licenses/OSL-3.0

• Rishabh Iyer
• John Halloran
• Kai Wei

1. Convex Function API

• Base class for convex optimization
• L1LogistocLoss and L2LogistocLoss,
• L1SmoothSVMLoss and L2SmoothSVMLoss,
• L1HingeSVMLoss and L2HingeSVMLoss,
• L1ProbitLossLoss and L2ProbitLoss,
• L1HuberSVMLoss and L2HuberSVMLoss,
• L1SmoothSVRLoss and L2SmoothSVRLoss,
• L1HingeSVMLoss and L2HingeSVMLoss

2. Convex Optimization Algorithms API

• Trust Region Newton (TRON)
• LBFGS Algorithm
• LBFGS OWL (L1 regularization)
• Conjugate Gradient Descent
• Dual Coordinate Descent for SVMs (SVCDual)
• Gradient Descent
• Gradient Descent with Line Search
• Gradient Descent with Nesterov's algorithm
• Gradient Descent with Barzilai-Borwein step size
• Stochastic Gradient Descent
• Stochastic Gradient Descent with AdaGrad
• Stochastic Gradient Descent with Dual Averaging
• Stochastic Gradient Descent with Decaying Learning Rate

3. ML Classification API

• L1 Logistic Regression,
• L2 Logistic Regression
• L1 Smooth SVM
• L2 Smooth SVM
• L2 Smooth SVM

4. ML Regression API

• L1 Linear Regression
• L2 Linear Regression
• L1 Smooth SVRs
• L2 Smooth SVRs
• L2 Hinge SVRs

1. Install CMake
2. Go to the main directory of jensen
3. mkdir build
4. cd build/
5. cmake ..
6. make

Once you run make, it should automatically build the entire library. Once the library is built, please try out the example codes in the build directory.

To test the optimization algorithms please run the test executables: ./TestL1LogisticLoss ./TestL2LogisticLoss ./TestL1SmoothSVMLoss ./TestL2LeastSquaresLoss etc.

You can also play around with the examples for testing classification and regression models. You can try them out as: ./ClassificationExample -trainFeatureFile ../data/heart_scale.feat -trainLabelFile ../data/heart_scale.label -testFeatureFile ../data/heart_scale.feat -testLabelFile ../data/heart_scale.label Optionally you can also play around with the method (L1LR, L2LR etc.), the algtype (LBFGS, TRON etc.), the regularization and so on.