inst/coxprocess/demo_gibbsflow_ais_vi.R

In jeremyhengjm/GibbsFlow: Gibbs Flow for Approximate Transport

rm(list = ls())
library(GibbsFlow)
library(spatstat)
library(tictoc)
library(ggplot2)

# load pine saplings dataset
data(finpines)
data_x <- (finpines$x + 5) / 10 # normalize data to unit square
data_y <- (finpines$y + 8) / 10
plot(x = data_x, y = data_y, type = "p")

ngrid <- 20
grid <- seq(from = 0, to = 1, length.out = ngrid+1)
dimension <- ngrid^2
data_counts <- rep(0, dimension)
for (i in 1:ngrid){
  for (j in 1:ngrid){
    logical_y <- (data_x > grid[i]) * (data_x < grid[i+1])
    logical_x <- (data_y > grid[j]) * (data_y < grid[j+1])
    data_counts[(i-1)*ngrid + j] <- sum(logical_y * logical_x)
  }
}

# artificial prior
prior <- list()
prior$logdensity <- function(x) as.numeric(coxprocess_log_vi_proposal(x))
prior$gradlogdensity <- function(x) coxprocess_gradlog_vi_proposal(x)
prior$rinit <- function(n) coxprocess_sample_vi_proposal(n)

# artificial likelihood
likelihood <- list()
likelihood$logdensity <- function(x) as.numeric(coxprocess_logprior(x)) + 
  as.numeric(coxprocess_loglikelihood(x, data_counts)) - 
  as.numeric(coxprocess_log_vi_proposal(x))
likelihood$gradlogdensity <- function(x) coxprocess_gradlogprior(x) + 
  coxprocess_gradloglikelihood(x, data_counts) - 
  coxprocess_gradlog_vi_proposal(x)

# define functions to compute gibbs flow (and optionally velocity)
exponent <- 2
compute_gibbsflow <- function(stepsize, lambda, lambda_next, derivative_lambda, x, logdensity) coxprocess_gibbsflow_vi_proposal(stepsize, lambda, derivative_lambda, x, logdensity, data_counts)
gibbsvelocity <- function(t, x) as.matrix(coxprocess_gibbsvelocity_vi_proposal(t, x, exponent, data_counts))

# smc settings
nparticles <- 2^9
nsteps <- 80
timegrid <- seq(0, 1, length.out = nsteps)
lambda <- timegrid^exponent
derivative_lambda <- exponent * timegrid^(exponent - 1)

# mcmc settings
mcmc <- list()
mcmc$choice <- "rmhmc"
mcmc$parameters$stepsize <- 0.20
mcmc$parameters$nsteps <- 10
mcmc$nmoves <- 1

# compute metric tensor as in Girolami and Calderhead 2011 (Section 9)
parameter_sigmasq <- 1.91
parameter_mu <- log(126) - 0.5 * parameter_sigmasq
parameter_beta <- 1 / 33
parameter_area <- 1 / dimension
prior_cov <- matrix(nrow = dimension, ncol = dimension)
for (m in 1:dimension){
  for (n in 1:dimension){
    index_m <- c( floor((m-1) / ngrid) + 1, ((m-1) %% ngrid) + 1 )
    index_n <- c( floor((n-1) / ngrid) + 1, ((n-1) %% ngrid) + 1 )
    prior_cov[m,n] <- parameter_sigmasq * exp(- sqrt(sum((index_m - index_n)^2)) / (ngrid * parameter_beta) )
  }
}
prior_precision <- solve(prior_cov)
metric_tensor <- prior_precision
diag(metric_tensor) <- parameter_area * exp(parameter_mu + 0.5 * diag(prior_cov)) + diag(prior_precision)
mcmc$parameters$metric$inverse  <- solve(metric_tensor)
metric_chol_inverse <- t(chol(mcmc$parameters$metric$inverse))
mcmc$parameters$metric$inverse_chol_inverse <- solve(t(metric_chol_inverse))

# free up memory
prior_cov <- NULL
prior_precision <- NULL
metric_tensor <- NULL
metric_chol_inverse <- NULL

# run ais/smc
tic()
# smc <- run_gibbsflow_ais(prior, likelihood, nparticles, timegrid, lambda, derivative_lambda, compute_gibbsflow, mcmc)
smc <- run_gibbsflow_ais(prior, likelihood, nparticles, timegrid, lambda, derivative_lambda, compute_gibbsflow, mcmc, gibbsvelocity)
toc()

# ess plot
ess.df <- data.frame(time = 1:nsteps, ess = smc$ess * 100 / nparticles)
ggplot(ess.df, aes(x = time, y = ess)) + geom_line() + 
  labs(x = "time", y = "ESS%") + ylim(c(0, 100))
smc$log_normconst[nsteps]

# normalizing constant plot
normconst.df <- data.frame(time = 1:nsteps, normconst = smc$log_normconst)
ggplot() + geom_line(data = normconst.df, aes(x = time, y = normconst), colour = "blue") + 
  labs(x = "time", y = "log normalizing constant")

# acceptance probability
acceptprob_min.df <- data.frame(time = 1:nsteps, acceptprob = smc$acceptprob[1, ])
acceptprob_max.df <- data.frame(time = 1:nsteps, acceptprob = smc$acceptprob[2, ])
ggplot() + geom_line(data = acceptprob_min.df, aes(x = time, y = acceptprob), colour = "blue") + 
  geom_line(data = acceptprob_max.df, aes(x = time, y = acceptprob), colour = "red") + ylim(c(0, 1))

# norm of gibbs velocity
normvelocity.df <- data.frame(time = timegrid, 
                              lower = apply(smc$normvelocity, 2, function(x) quantile(x, probs = 0.25)),
                              median = apply(smc$normvelocity, 2, median),
                              upper = apply(smc$normvelocity, 2, function(x) quantile(x, probs = 0.75)))
gnormvelocity <- ggplot(normvelocity.df, aes(x = time, y = median, ymin = lower, ymax = upper))
gnormvelocity <- gnormvelocity + geom_pointrange(alpha = 0.5) + 
  xlim(0, 1) + # scale_y_continuous(breaks = c(0, 40, 80, 120)) +
  xlab("time") + ylab("norm of Gibbs velocity") 
gnormvelocity
save(normvelocity.df, file = "inst/coxprocess/results/normvelocity_gfais_vi_400.RData")
# ggsave(filename = "~/Dropbox/GibbsFlow/draft_v3/coxprocess_normvelocity_gfais_vi_400.pdf", plot = gnormvelocity,
# device = "pdf", width = 6, height = 6)