Miguel Lazaro-Gredilla Homepage
Home CV Publications Downloads MLG

Advanced usage of the SSGP code

It is recommended that you first read the simple usage tutorial. Refer to the paper for details of the algorithm.

The files

Once you have downloaded and unzipped the code, you should have five files:
  1. ssgpr_ui.m - Main function you should be calling. It handles the initialization of hyperparameters as described in the paper.
  2. ssgprfixed_ui.m - Same for fixed spectral points.
  3. ssgpr.m - Code for SSGP regression. It provides training Negative Log Marginal Likelihood (NLML), derivatives with respect to the hyperparameters and predictive means and variances for test data.
  4. ssgprfixed.m - Same for fixed spectral points.
  5. minimize.m - Generic multivariate function optimization (written by Carl E. Rasmussen).

Using ssgprfixed_ui.m to perform regression

Syntax

[NMSE, mu, S2, NMLP, loghyper, convergence] = ssgprfixed_ui(x_tr, y_tr, x_tst, y_tst, m, iteropt, loghyper)

Returns:

Examples

As you see, you can do almost anything you want with the flexibility provided by this interface.

Using ssgpr_ui.m to perform regression

Syntax

[NMSE, mu, S2, NMLP, loghyper, convergence] = ssgpr_ui(x_tr, y_tr, x_tst, y_tst, m, iteropt, loghyper)

One step further: Calling directly the main SSGP code

If you are writing your own specialized code, you may want to directly access the core code, ssgprfixed (or equivalently, ssgpr). In most cases, though, you will want to avoid this, and use the convenient interface provided by ssgprfixed_ui.

Syntax (for training)

[NLML, deriv_NLML] = ssgprfixed(hyper, x_tr, y_tr)

Returns:

Syntax (for prediction)

[mu, S2] = ssgprfixed(hyper, x_tr, y_tr, x_tst)

Returns:

© 
Miguel Lázaro-Gredilla
Last modified: 2009-03-05, 17:26