[r] Rewrite LSI/binned feature selection to utilize new normalization functions, add DimReduction S3 Class#181
Conversation
immanuelazn
changed the title
| #' @field cell_embeddings (IterableMatrix, dgCMatrix, matrix) The projected data | ||
| #' @field fitted_params (list) A list of parameters used for the transformation of a matrix. | ||
| #' @export | ||
| DimReduction <- function(x, fitted_params = list(), ...) { |
There was a problem hiding this comment.
I think this is the part that I am most iffy about. I don't see much in the codebase regarding new S3 classes, so I did not have much to reference off of. I am also iffy about how we want to expose the documentation to users. I created a new DimReduction field in pkgdown.yml to store this and project(). Any feedback on this section specifically would be greatly appreciated!
| #' @return IterableMatrix object of the projected data. | ||
| #' @details DimReduction subclasses should use this to project new data with the same features, to project into the same latent space. | ||
| #' All required information to run a projection should be held in x$fitted_params, including pertinent parameters when construction the DimReduction subclass object. | ||
| #' If there are no rownames, assume that the matrix is in the same order as the original matrix, assuming that they have the same number of features. |
There was a problem hiding this comment.
The row matching logic intuitively makes sense to me to have, but I recognize this wasn't in the design docs. I was able to accomplish this in LSI() by providing an additional attr named feature_names. Should this be reflected in the DimReduction base class as well?
| } | ||
|
|
||
|
|
||
| #' @export |
There was a problem hiding this comment.
Should I put a docstring for this or is this intuitive enough for S3 dispatch?
…`save_lsi`, remove `save_in_memory` in `lsi()`
1 participant
Details
This PR largely builds upon the previous PR from #156 for the LSI implementation. However, there are some primary changes to make it fit within the
DimReductionframework noted in the design docs on #167.Firstly an implementation of an S3 class
DimReductionwas made. DimReduction returns the results of a dimensionality reduction in attrcell_embeddings. Another attribute namedfitted_paramsprovides all the information required to re-run a dimensionality reduction, keeping all the fittings from the original data passed to create the class.fitted_paramsenables theprojectgeneric method, and can be built upon by all subclasses ofDimReduction.LSI implements both a subclass of
DimReductionand provides aprojectmethod as well. The main fitted information utilized by a new run ofprojectis the feature means when doing normalization, and the SVD params.Tests are the same as in the previous PR, except with changes in syntax to reflect the new object styling