The PermonSVM package provides an implementation of binary classification via soft-margin Support Vector Machines (SVM), designed for use on clusters. PermonSVM implements scalable training procedure based on a linear kernel, taking advantage of an implicit representation of the Gramm matrix, and scalable matrix vector product of PETSc matrices. The resulting quadratic programming (QP) problem with an implicit Hessian is solved by scalable QP solvers within the PermonQP package, combining QP transforms, an augmented Lagrangian approach (SMALXE), and efficient solvers for box constrained QP.
Our implementation of the SVM training procedure offers efficient utilization of parallel computers up to thousands of processor cores, and allows to substantially shorten the time of SVM training.
The largest problem successfully solved using the PermonSVM was the benchmark of suspicious URL prediction with more than 2 million examples and over 3 million features. Reached accuracy was 99.09% computed on 120 cores in 30 seconds.
Another real-world problem solved using our package is a ground truth learning in material engineering, it is detecting the brittle and ductile fractures on the DWTT specimen surface.
- scalable solution of binary classification SVM problems
- efficient implementation of the linear kernel
- L1-norm and L2-norm loss function
- parser for LIBSVM-format datasets
- fast, load-balanced cross-validation/grid search
- easy-to use PETSc-like API
- dependencies: PermonQP and PETSc 3.6 or higher
Public git repository: https://github.com/it4innovations/permonsvm.
Please use this site to send us issues and pull requests.