jlmelville / smallvis

R package for dimensionality reduction of small datasets

Writing

I think about dimensionality reduction a lot, but currently have no need or urge to publish in the stick-a-PDF-on-arXiv sense. As a result, I have managed to accumulate various markdown documents and notebooks scattered across many repositiories. I am losing track of it all myself, so here I collect in one place some of the discussions, experiments and other musings.

• A comparison of the cost functions and gradients of t-SNE, LargeVis and UMAP: https://jlmelville.github.io/smallvis/theory.html and https://jlmelville.github.io/uwot/articles/umap-for-tsne.html.
• LEOPOLD: an approximation to densMAP available in uwot: https://jlmelville.github.io/uwot/articles/leopold.html.
• Some recent explorations of dimensionality reduction, validation and visualization in Python using Jupyter notebooks: https://github.com/jlmelville/drnb/tree/master/notebooks.
• Some Python notebooks that grab a variety of datasets that you might either enjoy running dimensionality reduction methods on, or just useful example code of how to use Python libraries to do some initial data reading and processing: https://github.com/jlmelville/drnb/tree/master/notebooks/data-pipeline. A more limited R package: https://github.com/jlmelville/snedata.
• Graph Laplacians, Laplacian Eigenmaps, Diffusion maps and other spectral methods: https://jlmelville.github.io/smallvis/spectral.html.
• A discussion of using my nearest neighbor descent package to diagnose hubness in datasets and their affect on neighbor retrieval: https://jlmelville.github.io/rnndescent/articles/hubness.html.
• The smallvis doc page collects some of the above, plus discusses many other aspects of dimensionality reduction, especially t-SNE and variants: https://jlmelville.github.io/smallvis/.
• How to derive the gradient of neighbor-embedding methods: http://jlmelville.github.io/sneer/gradients.html.
• A (currently very out of date) bibliography of dimensionality reduction papers: http://jlmelville.github.io/sneer/references.html.
• The different forms of Nesterov momentum as used in stochastic gradient descent methods: https://jlmelville.github.io/mize/nesterov.html. Nothing much to do with dimensionality reduction, but you have to optimize those embeddings somehow.

smallvis's Issues

Error in if (Z == 0) { : missing value where TRUE/FALSE needed

Thank you for such an extensive package and for including examples for each and every function. The example with iris runs no problem; I try to reproduce for diff dataset of similar str() but I get the following error:

Load dataset, inspect str

i:

URL  <- "http://bit.ly/data_shannon_error"
goblins_df <- read.csv(URL) 
utils::str(goblins_df)

o:

'data.frame':	373 obs. of  9 variables:
 $ bone_length  : num  0.355 0.324 0.63 0.472 0.469 ...
 $ rotting_flesh: num  0.351 0.431 0.581 0.494 0.517 ...
 $ hair_length  : num  0.466 0.205 0.647 0.557 0.621 ...
 $ has_soul     : num  0.781 0.258 0.505 0.564 0.625 ...
 $ red          : int  255 255 255 255 255 255 255 255 255 255 ...
 $ green        : int  255 255 255 255 255 255 255 255 255 255 ...
 $ blue         : int  255 255 255 255 255 255 255 255 255 255 ...
 $ alpha        : int  0 0 0 0 0 0 0 0 0 0 ...
 $ type         : Factor w/ 3 levels "Ghost","Ghoul",..: 2 1 2 2 3 2 2 2 3 3 ...

smallvis() call

i:

tsne <- smallvis(data, 
                          perplexity = 30, 
                          verbose    = TRUE)

Error:

o:

Error in if (Z == 0) { : missing value where TRUE/FALSE needed
6.shannon(Di, beta[i])
5.x2aff(X, perplexity, tol = 1e-05, kernel = kernel, verbose = verbose)
4.sne_init(cost, X, perplexity = perplexity, kernel = inp_kernel, symmetrize = "symmetric", normalize = TRUE, verbose = verbose, ret_extra = ret_extra)
3.cost$init(cost, X, verbose = verbose, ret_extra = ret_extra, max_iter = max_iter)
2.cost_init(cost_fn, X, max_iter = max_iter, verbose = verbose, ret_extra = ret_optionals)
1.smallvis(data, perplexity = 30, verbose = TRUE)

Does it have sth to do with shannon()? I ctrl + F in your code and source code from pgirmess ::shannon() but can't seem to detect where this error is exported from.

Unable to install smallvis

Hi there, I followed your readme and installed RSpectra and vizier successfully, but could not install smallvis. Hopefully there is something small I missed!

Thank you.

> library(RSpectra)
> library(vizier)
> devtools::install_github("jlmelville/smallvis/smallvis")
Downloading GitHub repo jlmelville/smallvis@HEAD
Error in utils::download.file(url, path, method = method, quiet = quiet,  : 
  download from 'https://api.github.com/repos/jlmelville/smallvis/tarball/HEAD' failed
> R.version.string
[1] "R version 4.2.1 (2022-06-23)"
> sessionInfo()
R version 4.2.1 (2022-06-23)
Platform: x86_64-apple-darwin17.0 (64-bit)
Running under: macOS Monterey 12.6.1

Matrix products: default
LAPACK: /Library/Frameworks/R.framework/Versions/4.2/Resources/lib/libRlapack.dylib

locale:
[1] en_GB.UTF-8/en_GB.UTF-8/en_GB.UTF-8/C/en_GB.UTF-8/en_GB.UTF-8

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

other attached packages:
 [1] devtools_2.4.5        usethis_2.1.6         vizier_0.4.2          RSpectra_0.16-1       data.table_1.14.6     ComplexHeatmap_2.12.1
 [7] BiocParallel_1.30.4   forcats_0.5.2         stringr_1.4.1         dplyr_1.0.10          purrr_0.3.5           readr_2.1.3          
[13] tidyr_1.2.1           tibble_3.1.8          ggplot2_3.4.0         tidyverse_1.3.2       SeuratObject_4.1.3    Seurat_4.3.0         
[19] SCopeLoomR_0.13.0     SCENIC_1.3.1          AUCell_1.18.1         flowCore_2.8.0       

A reference for scaled PCA initialisation

Hey, do you by any chance know of any reference for the scaled PCA initialization? Perhaps there is some paper that at least mentions PCA initialization as a possibility and explicitly says that one needs to use scaled PCA in this case? I am not able to find any reference at all.

I looked here https://jlmelville.github.io/smallvis/init.html and here http://jlmelville.github.io/sneer/references.html but did not find anything. In fact, in your References list you cite our preprint in the Initialization setting, but I think I borrowed the scaled PCA idea from your pages :-) I am preparing a revision right now and it'd be good to have some reference for this...

smallvis_rep() not working

Very useful and fun package. But trying to run the sample code, I get an error messge.

library(smallvis)
tsne_iris_best <- smallvis_rep(nrep = 10, iris, perplexity = 25, ret_extra = TRUE)

23:16:10 Starting embedding # 1 of 10
Error in make_smallvis_cb(X) : argument "X" is missing, with no default

Thanks!