gpflow: GPflow Python Module
In goldingn/gpflowr: Gaussian Process Modelling Using TensorFlow

Description Usage Details Documentation Examples

gpflow is the imported GPflow python module. You can access the various methods via the $ operator. For example, to access the submodule of kernel constructor functions type gpflow$kernels.

gpflow_available() tests whether the GPflow module is installed and available to the version of Python being used by reticulate.

1	gpflow_available()

reticulate converts R objects to python objects on the fly. It won't always guess how to do that correctly though, so here are a few gotchas to look out for:

data - GPflow expects vectors to be *column* vectors. Vectors should there be converted to one-column matrices before being passed to most functions. E.g. using matrix(), as in the example below.
integers - Python (and therefore GPflow) is more stringent about integers than R is. For example, when specifying the number of active dimensions of a kernel, you should make sure the number passed is of type integer, not numeric. In most cases that just means adding an 'L', i.e. switching 1 to 1L etc.

If you want to tidy up your code, you can rename gpflow by assigning it to a new object (e.g. gp <- gpflow).

The submodules and methods in gpflow are not documented in this package. You can access the python docs at http://gpflow.readthedocs.io/en/latest/.

If you're using a recent version of the Rstudio editor, you can also get inline help and autocompletion by hitting the tab key.

if (gpflow_available()) {

  # translation of the simple GP regression example from GPflow:
  # http://gpflow.readthedocs.io/en/latest/notebooks/regression.html
  
  # fake data
  N <- 12
  X <- matrix(runif(N))
  Y = sin(12 * X) + 0.66 * cos(25 * X) + matrix(rnorm(N)) * 0.1 + 3
  plot(X, Y)
  
  # set up kernel and GP
  k <- gpflow$kernels$Matern52(1L, lengthscales = 0.3)
  m <- gpflow$gpr$GPR(X, Y, kern=k)
  
  # adjust the likelihood
  m$likelihood$variance = 0.01
  
  # predict and plot
  xx <- matrix(seq(-0.1, 1.1, len = 100))
  tmp <- m$predict_y(xx)
  mean <- tmp[[1]]
  lines(mean ~ xx)
 
}