Deprecations

• require_monoid= keyword in ewise_add has been deprecated (#236)
  • New behavior is like require_monoid=False and will no longer raise

Improvements

• Add IndexUnaryOp and gb.indexunary namespace (#229)
  • Apply works with both IndexUnaryOp and SelectOp
• Add UDF support for IndexUnaryOp and SelectOp (#233)
• Add x.get method to Scalar, Vector, and Matrix objects, with optional default value (#241)
  • These return a Python scalar of the element at the location, or default value if element is missing
• Make aggregators callable (#242)
  • agg.sum(A) same as A.reduce_scalar(agg.sum)
  • agg.max(A, rowwise=True) same as A.reduce_rowwise(agg.max)
• Add mask.new(), which can also merge two masks efficiently (#240)
  • The result is True wherever the mask(s) would have been applied, and empty otherwise
  • complement=True returns the complement
• Allow A << 1 to create dense (iso-valued) matrices (this used to raise) (#243)
• Add “index” and “value” strings to Vector repr (#226)
• Add cbrt from SuiteSparse:GraphBLAS 7.1.0 to compute principal cube root (#244)
  • For example, cbrt(-8) == -2, and the result is a float
  • Also, add numpy ufuncs cbrt, float_power, and positive to the numpy namespaces (#237)
• op.from_string now includes aggregators (last priority) (#235)
  • e.g., op.from_string("count") now works
• Allow record dtypes to be created via dicts, such as {'x': int, 'y': float} (#227)
• Allow scalars with record dtypes to be created from dicts (#228)
  • e.g., Scalar.from_value({'x': 1, 'y': True}, dtype=udt)
• Added scripts/ folder in repo to help developers/contributors perform useful tasks (#231)

• Improvements to operators and user-defined functions (UDFs):
  • Allow Python functions to be passed as BinaryOp or UnaryOp, which will be automatically jitted
  • Add register_new and register_anonymous functions to e.g. gb.binary and gb.semiring namespaces
  • Add binary.binom to compute binomial coefficients exactly
• Improvements to dtypes and user-defined types (UDTs):
  • Auto-register UDTs, which means non-builtin NumPy dtypes may be passed as the dtype
  • Improved default names of UDTs
  • Improved dtype inference for first, second, and pair binary operators
    • For example, first uses the dtype of the first argument, and pair defaults to INT64
• Add .reposition(...) to Matrix and Value objects, which “shift” or reposition elements
• Broadcasting Vector to Matrix (such as plus(A & v)) has better repr and is more efficient with masks
• Add sort=True keyword to .to_vector(...) method on vectors and matrices
• Deprecate mask.mask and rename to mask.parent
• Added git pre-commit hooks to make it easier for contributors

Renamed project to Python-graphblas and package to graphblas (#204)

• $ conda install python-graphblas
• $ pip install python-graphblas
• >>> import graphblas as gb

Deprecations (#205)

• Deprecate how we create empty objects:
  • Scalar.new(dtype) -> Scalar(dtype)
  • Vector.new(dtype) -> Vector(dtype)
  • Matrix.new(dtype) -> Matrix(dtype)
  • Also, dtype argument is now optional (default FP64)

Features

• Support select! (#207)
  • gb.select.tril(A)
  • A.select("diag", 1)
  • gb.select.row(A < 3)
  • More to come; this is part 1 of the plan
• Support user-defined types! (#177)
  • gb.dtypes.register_new(name, dtype) (and register_anonymous)
  • Any NumPy dtype that Numba can handle in no-python mode is valid
    • For example, record dtypes and subarray dtypes
  • Many operations such as eq, first, positioni, and agg.count that you expect to work, do.
• Auto-broadcast vectors with diag and outer product (#203)
  • plus(A & v) -> any_plus(A @ v.diag())
  • and infix, such as A + v and v - A (using normal infix rules)
• Support ewise-union in infix (#203)
  • plus(x | y, left_default=1, right_default=2)
• Add x.ss.iter{keys,values,items} (#201)
  • This exposes fast SuiteSparse:GraphBLAS iteration
  • The object should not be mutated during iteration

Bug fixes

• Fixed segfault from using complex scalars and complex monoids (#164, #210)
• Misc

Deprecation cycle! (#190)

• Removed:
  • matrix.reduce_rows (use matrix.reduce_rowwise instead)
  • matrix.reduce_columns (use matrix.reduce_columnwise instead)
  • matrix.ss.scan_rows (use matrix.ss.scan_rowwise instead)
  • matrix.ss.scan_columns (use matrix.ss.scan_columnwise instead)
  • grblas.backends removed
• Deprecate:
  • vector.ss.diag(A) (renamed to vector.ss.build_diag(A); use A.diag instead)
  • matrix.ss.diag(v) (renamed to matrix.ss.build_diag(v); use v.diag instead)

Other changes

• Autocompute scalar expressions in x.update(expr) if valid (#192)
• Improve tab-completion of operator namespaces (#191)
• Add missing semirings that were the commutation of existing semirings (#191)
  • This added semirings with rpow, rtruediv, rfloordiv, and abssecond binary operators.
• Update v += expr to do v(plus) << expr instead of v << plus(v | expr) when expr is not scalar-like (#193)
• Add x.ss.iso_value to Vector and Matrix objects (#194)
• Improved gb.io.from_networkx (#195)
  • Expose arguments to networkx.to_scipy_sparse_array
  • Use Matrix.ss.import_csr for faster importing of data when using “suitesparse” backend
  • Determine whether the matrix is iso-valued when using “suitesparse” backend

• Add gb.agg.from_string
• Make diag a mainline feature (still available under .ss namespace)
• Fix some failing tests

Lot’s of changes this release!

• Switch versioning to use Calendar versioning (calver)
• Support (and test on) Python 3.10
• Update to SuiteSparse:GraphBLAS 6.2.X
• Breaking change: reductions to scalar on empty collections now result in empty scalars by default (#162)
  • This affects my_vector.reduce and my_matrix.reduce_scalar
  • To get old behavior where the monoid identity is returned, do e.g. v.reduce(allow_empty=False).
• Add support for GrB_Scalar objects (#163)
  • A Scalar object in grblas may be a C scalar or GrB_Scalar (the default).
  • Use is_cscalar=True upon creating the Scalar to get a C scalar
  • Scalars may now be passed to gb.ss.concat as length-1 vectors or shape (1, 1) matrices
  • Scalars now have c.is_cscalar and c.is_grbscalar attributes
• Add ewise_union method (#159)
  • This is like ewise_add with a Binary Operator and user-provided defaults for the left and right values.
  • Subtraction infix x - y on Matrices or Vectors now use ewise_union with default values of 0.
• Add .ss.compactify to Vector and Matrix objects (144)
  • For example, this can shift all (or N) values contiguously to the left.
  • Supported operations: first, last, smallest, largest, random.
• Add .is_positional attribute to operators (#161)
  • For example, unary.positioni, binary.firsti, and semiring.any_secondj are “positional”
• Autocompute A[idx] as extract (#166)
  • Also, add better reprs for indexed expressions
• Support deleting many items using slice or indexing notation (#163)
  • For example, del A[[1, 2], :3]
• Support negative indices (#143)
• Allow resolved indices to be converted back to Python indices (#142)
• Add __sizeof__ method to Matrix, Vector, and Scalar objects to get memory usage ($145)
• Added a code of conduct (#167)
  • See: https://github.com/metagraph-dev/grblas/blob/main/CODE_OF_CONDUCT.md

Fix failing import in Python 3.10

• Update to SuiteSparse:GraphBLAS 6.1.3
• Add format such as "csr" to object reprs
• Much faster import time!
• All grblas objects may now be pickled/serialized
• grblas.ss.concat can now concatenate Vectors:
  • gb.ss.concat([v, v]) # Vector, size=2*N
  • gb.ss.concat([[v], [v]]) # Matrix, nrows=2*N, ncols=1
  • gb.ss.concat([[v, v]]) # Matrix, nrows=N, ncols=2
• Add Vector.ss.split to split a Vector into sub-Vectors
• Add .ndim attribute to Scalar, Vector, and Matrix objects and expressions

• Support infix between non-scalar (i.e., Matrix and Vector) objects! Here are the rules:
  • monoids with identity 0 use ewise_add: +,^, and |.
  • A - B will use GxB_eWiseUnion with 0 as the defaults (not yet supported!)
  • others use ewise_mult
• Use numpy rules for unifying types, including when scalars are used.
  • For example, A + c with A.dtype == 'INT8' and c.dtype == 'INT64' will result in INT8 dtype by default.
  • See: https://numpy.org/doc/stable/reference/generated/numpy.find_common_type.html
• Allow ss.concat to use expressions, such as ss.concat([[A, A.T, A + 1]])
• Fix: removed op.div, added op.truediv, and made sure op namespace matches other operator namespaces

Deprecations:

• Changed e.g. reduce_rows to reduce_rowwise and deprecate the former

Updates:

• Added x.ss.selectk(how, k) methods for vectors and matrices
  • how can be "random", "first", "last". Vector also supports "largest" and "smallest"
• Allow operators to be given as strings, such as A.mxm(B, 'min.+[float]')
  • Also added from_string functions such as gb.binary.from_string('+')
• x | y, x & y, x ^ y, and ~x infix is only value-y for bool dtypes
  • x | y and x ^ y uses ewise_add, x & y uses ewise_mult
• Add op.commutes_to and op.is_commutative to binaryops, monoids, and semirings
• Support slices from SuiteSparse (can now perform huge slices on huge objects!)
• Recorder now logs exceptional C calls and their errors
• Improved get_semiring(monoid, binaryop) to better handle ops from numpy namespaces.
• Make dtype= an optional positional argument (like NumPy does), not keyword-only
• Map numpy ops in e.g. gb.binary.numpy to GraphBLAS ops (by default)

Fixes:

• Fixed truediv on complex dtypes
• Fixed displaying of nan in reprs
• Fixed repr of large iso-valued objects
• Fixed types of land, lor, and lxor binary operators (not bool only)