greta blog

greta version 0.5.0

Nicholas Tierney — Thu, 29 Sep 2022 00:00:00 GMT

We are happy to say that greta version 0.5.0 is on CRAN!

This particular release has been a long time in the making, as it uses tensorflow version 2. We’ll refer to Tensor Flow as TF hereafter. The overall impact on greta users should be very minimal, but internally we have made substantial changes, and we wanted an opportunity to discuss the reasons why we needed to do this, and what this means for greta going forward.

TF version 2

Google released TF version 2 in … 2019. This was a big change in the user interface to TF, as it changed from what was called “graph mode” to “eager mode”. Briefly, eager mode is ..

What changed for greta

A lot. And also, not much. Internally we had to change a very large amount of the code insidie of greta, as changing from graph based execution to eager execution meant that

New features in greta

Upcoming features in greta

Now that we’ve implemented TF2 in greta, we can start taking advantage of the new features inside TF and implement them in greta. Here’s a list of some of the new features you can expect in the future

No U-turn Sampler
Faster Tuning
Marginalisation interface

greta version 0.4.3

Nicholas Tierney — Mon, 12 Sep 2022 00:00:00 GMT

We are happy to say that greta version 0.4.3 is now on CRAN

greta.dynamics 0.2.0

Nicholas Tierney — Tue, 06 Sep 2022 00:00:00 GMT

greta.gp 0.2.0

Nicholas Tierney — Tue, 06 Sep 2022 00:00:00 GMT

Bridges in Perth, film, Nick Tierney

We are excited to announce that greta.gp version 0.2.0 is now on CRAN!

greta.gp provides a syntax to create and combine Gaussian process kernels in greta. You can then them to define either full rank or sparse Gaussian processes.

Here’s a brief example:

# simulate data
x <- runif(20, 0, 10)
y <- sin(x) + rnorm(20, 0, 0.5)
x_plot <- seq(-1, 11, length.out = 200)

library(greta)
library(greta.gp)
# hyperparameters
rbf_var <- lognormal(0, 1)

rbf_len <- lognormal(0, 1)
obs_sd <- lognormal(0, 1)
# kernel & GP
kernel <- rbf(rbf_len, rbf_var) + bias(1)
f <- gp(x, kernel)
# likelihood
distribution(y) <- normal(f, obs_sd)
# prediction
f_plot <- project(f, x_plot)

# fit the model by Hamiltonian Monte Carlo
m <- model(f_plot)
draws <- mcmc(m, n_samples = 250)

# plot 200 posterior samples
plot(
  y ~ x,
  pch = 16,
  col = grey(0.4),
  xlim = c(0, 10),
  ylim = c(-2.5, 2.5)
)
for (i in 1:200) {
  lines(
    draws[[1]][i, ] ~ x_plot,
    lwd = 2,
    col = rgb(0.7, 0.1, 0.4, 0.1)
  )
}