Changes in version 0.1.1                        

bug fixes

  - A cluster label with no spots in the expression no longer stops the
    run with "quadratic program could not be solved". Named clusters may
    now be a superset, e.g. the clusters of a whole dataset when expr is
    a subset.
  - In a cluster with a single candidate cell type, a spot with no
    signal on the signature features is now NA, as it is in every other
    cluster; it was returned as proportion 1 of that type.
  - dwls_stream() matches the second round's spots to the first round's
    by name, so a read_window() that returns a window's spots in a
    different order on a later pass no longer gives silently wrong
    results.
  - Duplicated or missing cell type names, feature names or spot IDs are
    now an error; a duplicated cell type silently overwrote another's
    proportions. NA clusters get an error that says so.
  - Parallel runs no longer send every task's expression slice to each
    worker.

changes

  - Noted for 0.1.0: the PAGE step that picks candidate cell types
    reverse-logs with logbase. Giotto <= 4.3.1 used base 2 there
    whatever logbase was, so results differ from it slightly for logbase
    other than 2 (identical at the default).

                        Changes in version 0.1.0                        

First release as a standalone package. The algorithm was previously
runDWLSDeconv() inside Giotto (<= 4.3.1), which now calls this package.

Results match Giotto 4.3.1 to floating point (~1e-15), except where
these bugs were fixed:

  - Clusters with a single candidate cell type in the spot round now
    assign that type proportion 1; they were left at 0.
  - Small negative values from the constrained solver are now clipped
    to 0 for every cell type, not only the last one.
  - Running DWLS no longer leaves the session's random number generator
    in a fixed state.

Faster (12-22x serial on 600-20,000 spots):

  - The dampening constant's cross-validation uses .lm.fit() on k x k
    weighted least squares instead of lm(), and draws its 100 subsets
    once.
  - Weighted solves use crossprod(S * w, S) instead of a features x
    features diag(w), and call Matrix::nearPD() only when the matrix
    needs it.
  - The reverse-log transform keeps sparse input sparse and touches only
    the signature features.

New:

  - BPPARAM for parallel per-cluster plans and per-spot solves.
  - dwls_stream() for expression read in windows.
  - na_as_zero.
  - dwls_signature(), the matrix form of Giotto's
    makeSignMatrixDWLSfromMatrix().