dstructs / matrix Goto Github PK

Currently, matrices are row-major order. Supporting column-major order would require internal modifications to provide efficient access patterns.

• What are the use cases for supporting column-major order as an option?
• API implications?
  • Would probably require modifying the API from being variadic to being an object-based interface.
• How different from creating a Matrix and then just taking the transpose?

When using the non-raw implementation of the matrix module, it is not possible right now to assign the value NaN to a float32 or float64 matrix, even though the underlying typed arrays could naturally hold this value. This has been a problem for me in a couple of modules, as it conflicts with our standard of failing silently when a computation is not possible for the given input types. Your thoughts?

Currently, only accept arrays for methods such as mset. Should also accept typed-arrays, etc.

Currently, shape and strides are shallow copied. This is leaky and may lead to bugs due to (unintended) mutations.

The following modules should be used:

1. dstructs-cast-arrays to cast the internal typed array to a generic array
2. utils-copy to copy shape and strides

What are your thoughts on allowing for row and column names? R has this feature, and for standard applications in statistics, where the columns of a design matrix hold data for different predictors, this helps immensely. It allows oneself to index via the column and row names, for example.
Could we support something similar?

Here is an example:

mat = matrix( c( 40, 29, 55, 5000, 3900, 5500 ), nrow = 3, ncol = 2 )
colnames( mat ) = c(  "Age", "Income" )

Printing the matrix then gives the following:

    Age Income
[1,]  40   5000
[2,]  29   3900
[3,]  55   5500

The time has come.

NOTEs may be found in the source.