Get ready for BOSS 1.12, now with GPyTorch for GPUs

We have just updated BOSS with the latest features, including a GPy backend and optimisable noise!

  • to use non-standard GP models for multi-task, gradient observations etc a new keyword model_name has been introduced. This will default to single_task and does not need to be specified unless you want to use another type of model. Tutorials and documentation have been updated to reflect this change.
  • Added GPyTorch as a backend for GP models. This new backend can be activated by settings the new keyword model_backend=torch. If you have pytroch installed with cuda enabled, this will allow BOSS to run part of the calculations on the GPU, significantly reducing the calculation time for larger datasets. For now, only ordinary single task GP regression is supported.
  • Introduced a new keyword noise_optim that if set to True (default is False) will treat the likelihood noise as a hyperparameter to be optimized along with the kernel hyperparameters.
  • BOSS now prints the name of the current hyperparameters in a comment line to the boss.rst and hyperparamters.dat files.