Deep Compositional Spatial Models

Note: This version uses TensorFlowv2 -- for the original TensorFlowv1 version please download Release 0.2.0

Deep compositional spatial models are standard spatial covariance models coupled with an injective warping function of the spatial domain. The warping function is constructed through a composition of multiple elemental injective functions in a deep-learning framework. The package implements two cases for the univariate setting; first, when these warping functions are known up to some weights that need to be estimated, and, second, when the weights in each layer are random. In the multivariate setting only the former case is available. Estimation and inference is done using TensorFlow, which makes use of graphics processing units.

Resources

A manuscript detailing the theory and implementation in the univariate setting is available here, while a manuscript detailng the theory and implementation in a multivariate setting is available here. An informal blog post summarising the manuscript concerning the univariate setting is available here.

Installation Instructions

This is an R package. Please install devtools and then install this package by typing library("devtools") install_github("andrewzm/deepspat") in an R console.

Reproducible Code

Code using this package for reproducing the results shown in the manuscript describing the univariate setting is available in the supplemental material of our first article. Code for the results shown in manuscript describing the multivariate setting is available here. Please note that for this version of deepspat yyou will require R 3.6, TensorFlow 2.x and Python 3.7.10.

``` R version 3.6.3 (2020-02-29) Platform: x86_64-pc-linux-gnu (64-bit) Running under: Ubuntu 20.04.3 LTS

Matrix products: default BLAS: /usr/lib/x8664-linux-gnu/openblas-pthread/libblas.so.3 LAPACK: /usr/lib/x8664-linux-gnu/openblas-pthread/liblapack.so.3

locale: [1] LCCTYPE=enAU.UTF-8 LCNUMERIC=C
[3] LCTIME=enAU.UTF-8 LCCOLLATE=enAU.UTF-8
[5] LCMONETARY=enAU.UTF-8 LCMESSAGES=enAU.UTF-8
[7] LCPAPER=enAU.UTF-8 LCNAME=C
[9] LCADDRESS=C LCTELEPHONE=C
[11] LCMEASUREMENT=enAU.UTF-8 LC_IDENTIFICATION=C

attached base packages: [1] stats graphics grDevices utils datasets methods base

other attached packages: [1] dplyr1.0.5 deepspat0.1.2 deepGP0.1.0 testthat3.0.2 [5] jsonlite1.7.2 ggplot23.3.3 devtools2.4.0 usethis2.0.1
[9] reticulate_1.18

loaded via a namespace (and not attached): [1] fs1.5.0 xts0.12.1 FRK2.0.3
[4] RColorBrewer1.1-2 rprojroot2.0.2 tools3.6.3
[7] TMB1.7.20 backports1.2.1 utf81.2.1
[10] R62.5.0 rpart4.1-15 Hmisc4.5-0
[13] colorspace2.0-0 nnet7.3-13 withr2.4.2
[16] sp1.4-5 tidyselect1.1.0 gridExtra2.3
[19] prettyunits1.1.1 processx3.5.0 curl4.3
[22] compiler3.6.3 cli2.4.0 htmlTable2.1.0
[25] desc1.3.0 sparseinv0.1.3 scales1.1.1
[28] checkmate2.0.0 callr3.6.0 rappdirs0.3.3
[31] tfruns1.5.0 stringr1.4.0 digest0.6.27
[34] foreign0.8-75 rio0.5.26 base64enc0.1-3
[37] jpeg0.1-8.1 pkgconfig2.0.3 htmltools0.5.1.1
[40] sessioninfo1.1.1 fastmap1.1.0 readxl1.3.1
[43] htmlwidgets1.5.3 rlang0.4.10 rstudioapi0.13
[46] generics0.1.0 zoo1.8-9 tensorflow2.4.0
[49] zip2.1.1 car3.0-10 magrittr2.0.1
[52] Formula1.2-4 dotCall641.0-1 Matrix1.2-18
[55] Rcpp1.0.6 munsell0.5.0 fansi0.4.2
[58] abind1.4-5 lifecycle1.0.0 whisker0.4
[61] stringi1.5.3 carData3.0-4 plyr1.8.6
[64] pkgbuild1.2.0 grid3.6.3 parallel3.6.3
[67] forcats0.5.1 crayon1.4.1 lattice0.20-40
[70] haven2.4.1 splines3.6.3 hms1.0.0
[73] knitr1.33 ps1.6.0 pillar1.6.0
[76] ggpubr0.4.0 spacetime1.2-4 ggsignif0.6.1
[79] reshape21.4.4 pkgload1.2.1 glue1.4.2
[82] latticeExtra0.6-29 data.table1.14.2 remotes2.3.0
[85] png0.1-7 vctrs0.3.7 spam2.6-0
[88] cellranger1.1.0 gtable0.3.0 purrr0.3.4
[91] tidyr1.1.3 cachem1.0.4 xfun0.22
[94] openxlsx4.2.3 broom0.7.6 rstatix0.7.0
[97] survival3.1-8 tibble3.1.1 intervals0.15.2
[100] memoise2.0.0 cluster2.1.0 statmod1.4.35
[103] ellipsis_0.3.1
```

Note that when running within conda you will need to preface the R scripts with

library("reticulate")
use_condaenv("~/miniconda3/envs/TFv2/")

where you'd need to use your specific conda environment path.

I did not retain the exact original environment I used to produce the results but this environment should work. It was brought to my attention though that the DGPsparse code was not converging for the Monterrubio example with this environment. It does converge if you decrease the learning rates in the deepGP() function by a factor of ten in Experiment1D_DGPsparse.R:

learnrates = list(MH = 1e-7, SH = 1e-4, Z = 1e-9,
                  pars = par_learn_rate)

I also suggest fixing the seed in this script, so adding in the for loop of Experiment1D_DGPsparse.R:

set.seed(1)
tf$set_random_seed(1)
tf$random$set_random_seed(1)