search

davidsoergel / jlibsvm

Efficient training of Support Vector Machines in Java
Other
117 stars 62 forks source link

Race condition when classifying with MultiClassificationSVM #12

Open razmanolescu opened 6 years ago

razmanolescu commented 6 years ago

In MultiClassificationSVM::trainWithoutScaling there are parallel tasks spawned using Parallel.forEach . This restults in non-deterministic classification results. Serializing this fixes the problem.

The reason is that somehow the SVMs size is truncated and data is lost. My feeling is that the issue is within KernelQMatrix.