The KRR tools are split into two binaries, one for training and one for prediction. The data file consists of either explicit feature vectors or a kernel matrix, both of which should be in LIBSVM format.
krr-train [flags] data_file regularization_parameter [model_output_file]
Flags:
kernel
- The data_file
contains a kernel matrix, as oppose to feature vectors.
sparse
- Represent the feature vectors using a sparse data-structure.
dual
- Force the KRR problem to be solved in the dual.
primal
- Force the KRR problem to be solved in the primal.
approx
- Specify the rank to be used with a low-rank approximation of the kernel matrix (between 1 and # of training points).
krr-predict [flags] data_file model [predictions]
Flags:
kernel
- The data_file
contains a kernel matrix, as oppose to feature vectors.
sparse
- Represent the feature vectors using a sparse data-structure.
-- AfshinRostamizadeh - 11 Sep 2009