rgayler / vecsymr

Vector Symbolic Architecture (VSA) primitives for experimentation

vecsymr

The goal of {vecsymr} is to implement Vector Symbolic Architecture (VSA) primitives to support experimentation. It is intended to be a simple VSA implementation (the VSA equivalent of the geneticist’s fruit fly) to provide a convenient base for experimentation. The design choices are my personal preferences to support my research. The initial emphasis is on functionality and flexibility with no specific concern for performance.

I believe that phasor VSAs (where the vector elements are unit magnitude complex numbers) are the best choice for basic VSAs. However, I have initially imported functions for bipolar VSAs from VSA_altitude_hold to provide some code while I get the hang of writing an R package. Once the package contains enough phasor VSA code I will probably remove the bipolar VSA code. The phasor VSA code will probably include some extra features to support non-negativity and experimentation with clean-up memory.

Installation

You can install the development version of {vecsymr} from GitHub with:

# install.packages("devtools")
devtools::install_github("rgayler/vecsymr")

There is currently no intention to put this package on CRAN. If it turns out to be sufficiently useful and general I may try get it accepted as an rOpenSci package.

The current implementation is experimental. I expect the functional content to evolve as I work out what I want this package to do. I also expect the API to evolve as I work out how to make it easier to work with. If you want to do any reproducible work with the package you will need to use something like {renv} to freeze the version in use.

Ignore below here

Everything after this point is just boilerplate to be edited.

Example

This is a basic example which shows you how to solve a common problem:

library(vecsymr)
## basic example code

Remember

You’ll still need to render README.Rmd regularly, to keep README.md up-to-date. devtools::build_readme() is handy for this. You could also use GitHub Actions to re-render README.Rmd every time you push. An example workflow can be found here: https://github.com/r-lib/actions/tree/v1/examples.

If you create figures in the README don’t forget to commit and push the resulting figure files, so they display on GitHub.

vecsymr's Issues

Get CI R CMD check working

`testthat::expect_no_error()` should check the error text

expect_error will catch any error, not necessarily the {checkmate} errors I am looking for.
Look at using a regex to ensure the errors trapped come from {checkmate}
https://ropensci.slack.com/archives/C026GCWKC/p1665308114322969?thread_ts=1665305988.911269&cid=C026GCWKC

Look at automating the testing

The argument checking via {checkmate} is tediously repetitive to test.
Also, the VSA arguments to the operators will be all of the same "type".

Look at {patrick} and {autotest} as possible ways to simplify/automate testing.

https://ropensci.slack.com/archives/C026GCWKC/p1665368771849799?thread_ts=1665305988.911269&cid=C026GCWKC

https://ropensci.slack.com/archives/C026GCWKC/p1665382314171099?thread_ts=1665305988.911269&cid=C026GCWKC

`vsa_mk_atom_bipolar()` unit tests change the global environment

The vsa_mk_atom_bipolar() unit tests crank the random number generator - so alter .Random.seed
Look at using {withr} to cleanup after tests
https://ropensci.slack.com/archives/C026GCWKC/p1665308114322969?thread_ts=1665305988.911269&cid=C026GCWKC

Add `vsa_mk_atom_bipolar()` unit tests checking reproducibility

Check repeated calls with and without seed set to see that generated atoms are identical when intended.

Get CI test coverage badge working

Consider vectorised versions of hypervector functions

The currently defined (bipolar) functions have single hypervectors as the arguments and outputs.
Consider whether it would be useful to generalise these functions to work with collections (lists or vectors) of hypervectors.

The VSA_altitude_hold::run_simulation() function returns a data frame with list columns of hypervectors.

Torchhd generator functions return arrays of hypervectors (tensors), presumably fro compatibility with pytorch.

List columns of hypervectors are probably more idiomatic for R analyses, but it might eventually be useful for this package to be based on tensorflow for performance, in which case a tensor representation might be needed.