Reformat and reorganize for J. Data Science

blockedcholesky/.gitignore

@@ -0,0 +1,8 @@
+/.luarc.json
+/Manifest*.toml
+/jds.bst
+/jdsart.cls
+/*.ipynb
+/*.pdf
+/data/ml-32m.zip
+/data/ratings.arrow

blockedcholesky/Environment.qmd

@@ -0,0 +1,119 @@
+---
+title: "Computing environment and similar model fits in R"
+format: pdf
+author:
+  - name: Douglas Bates
+    email: dmbates@gmail.com
+    orcid: 0000-0001-8316-9503
+    affiliation:
+      - name: University of Wisconsin--Madison
+        city: University of Wisconsin--Madison
+        state: WI
+        url: https://www.wisc.edu
+        department: Department of Statistics
+  - name: Phillip M. Alday
+    email: phillip@phillipalday.com
+    orcid: 0000-0002-9984-5745
+    affiliation:
+      - name: Beacon Biosignals
+        city: Beacon Biosignals
+        url: https://beacon.bio
+  - name: Ajinkya H. Kokandakar
+    email: ajinkya@stat.wisc.edu
+    orcid: 0000-0001-6628-2272
+    affiliations:
+      - name: University of Wisconsin--Madison
+        city: University of Wisconsin--Madison
+        state: WI
+        url: https://www.wisc.edu
+        department: Department of Statistics
+monofont: JuliaMono
+date: last-modified
+number-sections: true
+execute:
+  cache: true
+engine: julia
+julia:
+  exeflags:
+    - -tauto
+    - --project=@.
+---
+
+# Computing environment
+
+The timings in the paper, with the exception of the fits of models to the `ml-32m` data, were performed on a laptop with the following characteristics
+
+```{julia}
+using AppleAccelerate, InteractiveUtils, LinearAlgebra, RCall
+versioninfo()
+```
+
+The versions of the Julia packages in use were
+
+```{julia}
+import Pkg
+Pkg.status()
+```
+
+and the accelerated BLAS used for this computer were
+
+```{julia}
+BLAS.get_config()
+```
+
+# Timing similar model fits with lme4
+
+For comparison we fit the same model as in section 3.1.1 using the `lme4::lmer` and `glmmTMB::tmbmod` functions in `R`
+
+```{r}
+dat <- lme4::InstEval
+dat$service <- as.numeric(dat$service) - 1  # convert to a 0/1 numeric vector
+ctrl <- lme4::lmerControl(calc.derivs = FALSE)
+form <- y ~ 1 + service + (1 | s) + (1 | d) + (1 | dept) + (0 + service | dept)
+mlmer <- lme4::lmer(form, dat, REML = FALSE, control = ctrl)
+print(summary(mlmer, correlation = FALSE))
+NULL       # suppress printing of the internal structure of the summary result
+```
+
+```{r}
+tmbmod <- glmmTMB::glmmTMB(form, dat)
+print(summary(tmbmod))
+NULL
+```
+
+These model fits are essentially the same as that from `MixedModels.jl` but they take longer
+
+```{r}
+#| warning: false
+print(
+  microbenchmark::microbenchmark(
+    lmer = lme4::lmer(form, dat, REML = FALSE, control = ctrl),
+    glmmTMB = glmmTMB::glmmTMB(form, dat),
+    times=6L,
+    unit="seconds"
+  )
+)
+NULL
+```
+
+Recall from section 3.1.1 of the paper that the `MixedModels.jl` fit of this model took just over one second on the same laptop computer.
+
+We attempted to fit the `ml-32m` example using `glmmTMB` on a desktop computer with 64 GiB of memory but shut it down after it had been running for more than a week on the machine without converging.
+Given otherwise similar performance, we do not expect `lme4` to perform any better than `glmmTMB`.
+
+# Versions of R and the R packages that were used
+
+```{r}
+print(R.version.string)
+print(packageVersion("lme4"))
+print(packageVersion("glmmTMB"))
+NULL
+```
+
+# Environment for the ml-32m model fits
+
+The environment for the ml-32m model fits is shown in the `config.txt` file.  The Julia source code to download the `ml-32m.zip` file, extract the `ratings` table, create `moviesnratings.arrow` and `usersnratings.arrow`, followed by `ratings.arrow` is `ml32-download.jl`.
+
+The Julia source file to run the `ml-32m` fits is `ml32-timefit.jl`.
+Before you start these fits remember that they take up to a day on a powerful server computer.
+Don't do this frivolously.

blockedcholesky/Makefile

@@ -7,7 +7,7 @@ preview:
 	quarto preview $(DOCNAME).qmd
 
 render:
-	quarto render $(DOCNAME).qmd --to jss-pdf+jss
+	quarto render $(DOCNAME).qmd --to jds-pdf
 
 render-arxiv:
 	sed -i'.bak' -e '/city:/d' $(DOCNAME).qmd

blockedcholesky/Project.toml

@@ -6,21 +6,26 @@ CSV = "336ed68f-0bac-5ca0-87d4-7b16caf5d00b"
 CairoMakie = "13f3f980-e62b-5c42-98c6-ff1f3baf88f0"
 Chairmarks = "0ca39b1e-fe0b-4e98-acfc-b1656634c4de"
 DataFrames = "a93c6f00-e57d-5684-b7b6-d8193f3e46c0"
+Downloads = "f43a241f-c20a-4ad4-852c-f6b1247861c6"
 MKL = "33e6dc65-8f57-5167-99aa-e5a354878fb2"
 MKL_jll = "856f044c-d86e-5d09-b602-aeab76dc8ba7"
 MixedModels = "ff71e718-51f3-5ec2-a782-8ffcbfa3c316"
+PRIMA = "0a7d04aa-8ac2-47b3-b7a7-9dbd6ad661ed"
 PrettyTables = "08abe8d2-0d0c-5749-adfa-8a2ac140af0d"
 Printf = "de0858da-6303-5e67-8744-51eddeeeb8d7"
+RCall = "6f49c342-dc21-5d91-9882-a32aef131414"
 SparseArrays = "2f01184e-e22b-5df5-ae63-d93ebab69eaf"
+Tables = "bd369af6-aec1-5ad0-b16a-f7cc5008161c"
+ZipFile = "a5390f91-8eb1-5f08-bee0-b1d1ffed6cea"
 
 [compat]
-AlgebraOfGraphics = "0.10.4"
+AlgebraOfGraphics = "0.11,0.12"
 Arrow = "2.8"
 CSV = "0.10"
-CairoMakie = "0.13.4"
+CairoMakie = "0.13,0.14,0.15"
 Chairmarks = "1.3"
 DataFrames = "1.7"
-MixedModels = "4.34"
-PrettyTables = "2.4"
+MixedModels = "4,5"
+PrettyTables = "2,3"
 Printf = "1"
-julia = "1.10"
+julia = "1.12"

blockedcholesky/README.md

@@ -3,26 +3,31 @@
 
 ## Steps
 
-- Install (or update) [`quarto`](https://quarto.org/) version 1.7 or greater (v1.6 may work as well, basically use any version of `quarto` that supports `engine: julia`)
+- Install the current Julia release version according to the instructions at https://julialang.org/downloads
+- Install (or update) [`quarto`](https://quarto.org/) version 1.9.0 or greater
 - Install `jupyter-cache` using the instructions from the quarto Julia docs [here](https://quarto.org/docs/computations/julia.html#jupyter-cache).
 - Install the `JuliaMono` font, see instructions at: https://juliamono.netlify.app
-- Install required `Julia` dependencies by running:
-   ```bash
-    $ julia --project=@. -e "import Pkg; Pkg.instantiate()"
-   ```
-- Generate the JSS formatted preview using:
+- Install the [`JDS extensions`](https://github.com/wenjie2wang/jds.qmd) for [`quarto`](https://quarto.org)
+- Install the required `Julia` dependencies by running:
     ```bash
-     $ make render
+    julia --project=@. -e "import Pkg; Pkg.instantiate()"
     ```
-- Generate the arXiv formatted version using:
+- Generate the JDL formatted preview using:
     ```bash
-    $ make render-arxiv
+    make render
     ```
-    **Note:** this command re-runs the `R` benchmark script which can take a long time.
-    **Note:** For the arxiv version, you need to have `cbfonts-fd` and `cbfonts` packages installed in your latex distribution
 - If you get errors due to caching, try using the `--no-cache` option for quarto:
     ```bash
-    quarto render BlockedCholeskyMM.qmd --no-cache                      # for JSS version render
-    quarto render BlockedCholeskyMM.qmd --no-cache --to arxiv-pdf+arxiv # for arxiv version
-
+    quarto render BlockedCholeskyMM.qmd --no-cache --to jds-pdf    # for JDS version render
+    ```
+- To create a Julia script from the Quarto file (`BlockedCholeskyMM.qmd`) convert the .qmd file to a Jupyter notebook and apply `jupyter nbconvert` to create a script.
+    ```bash
+    quarto convert BlockedCholeskyMM.qmd
+    jupyter nbconvert --no-prompt --to script BlockedCholeskyMM.ipynb
+    mv BlockedCholeskyMM.txt BlockedCholeskyMM.jl
     ```
+- Alternatively, `Quarto mode` for editors such as [VSCode](https://code.visualstudio.com) or [Positron](https://positron.posit.co) provide the ability to interactively evaluate code chunks from a Quarto document.
+
+- Comparison fits in R using the `lme4` and `glmmTMB` packages are shown in `Environment.pdf` generated from `Environment.qmd`
+
+- Because the ml-32m dataset cannot be redistributed, reproduction of the results in Table 4, require a separate script, `scripts/ml32-download.jl` to download the data, create the ratings table, and save it in the [Arrow](https://arrow.apache.org) format. The evaluation of the models is performed by `ml32-timefit.jl` **but** be aware that fitting very large models like this requires a computer with a large amount of memory (64 GiB is recommended) and can take a long time.  It took roughly 18 hours for all the model fits on a cloud compute instance with the characteristics shown in `config.txt`.

blockedcholesky/_extensions/wenjie2wang/jds/_extension.yml

@@ -0,0 +1,27 @@
+title: Journal of Data Science
+author: Wenjie Wang and Jun Yan
+version: 0.1.1
+quarto-required: ">= 1.4.0"
+contributes:
+  formats:
+    common:
+      shortcodes:
+        - shortcodes.lua
+    pdf:
+      documentclass: jdsart
+      classoption:
+        - letterpaper
+        - inpress
+      template: template.tex
+      cite-method: natbib
+      biblio-style: jds
+      biblio-config: false
+      fig-width: 6.5
+      fig-height: 4.5
+      format-resources:
+        - jdsart.cls
+        - jds.bst
+    html:
+      number-sections: true
+      toc: true
+      csl: apa.csl