1.11.0

This release adds official support for custom dataclasses in the phiml.dataclasses module.

Highlights

• Dataclass decorators @sliceable, @data_eq
• Replacing some fields of a dataclass instance keeps all unaffected caches from @cached_property
• Dataclasses no longer need to handle None values in comparisons when used in JIT or other tracing.
• Shorthand function *max, *min, *prod, *pack for specific dim types
• New syntax for item name renaming while slicing, e.g. tensor['y,x->a,b] selects the slices x and y and renames them.
• Add squeeze to remove singleton dims
• Improve formatting of sparse matrices
• clip() now accepts Shape for the upper limit
• Dim packing functions can now also be used for unpacking when passing one input and multiple output dims.

1.10.0

This release adds a lot of new functionality and pushes the new paradigm of letting users define custom types as @dataclass with @cached_property derived values.

• Expanded support for non-uniform tensors and sparse matrices
• Add tensor transpose shorthands Tensor.Ti, .Tc, .Ts and Tensor.dim.T. The old transpose is now called swap_axes.
• Add save, load to store using any (nested) tensor data using NumPy
• Add ncat and tcat variants (ccat, icat, scat, dcat)
• Add random_permutation, pick_random
• Functions such as slice, max, min now accept range to return the gather indices
• Add softmax, nan_to_0, counter_intersections
• Multi-Tensor unpacking in unpack
• Add Tensor.print
• Add *sum and *mean functions for the various dimension types
• Add map_d2b and map_c2d
• range is now also available as arange
• Deprecate rotation functions and move to PhiFlow
• Improved dataclass attribute detection, add experimental dataclass_getitem

Plus tons of bug fixes!

pip install phiml==1.10.0

1.9.3

Bug fixes plus

• Performance optimization for stack()
• experimental save/load
• at_min, at_max now support sparse matrices

1.9.2

Various smaller bug fixes, map now exposes expand_results.

1.9.0

Highlights

• convolve now behaves like matrix multiplication, reducing dual dims of the kernel
• Tensor @ Tensor can now be used to reduce channel dims in the absence of dual dims
• Improved support for shape spec strings, concat now supports packing using the syntax t->name:t
• Multi-dimensional cumulative_sum
• Improved support for non-uniform and sparse tensors
• New functions d2s, contains, count_occurrences, Tensor.map(), ravel_index and aliases rotate, cross
• Shape concatenation via Shape + Shape

1.8.0

Highlights

• NumPy 2 compatibility
• Tensor.numpy() and .native() now support dim packing
• wrap() and tensor() now support shape spec strings, e.g. 'example:b,(x,y,z)'
• Compact sparse tensors can now be created using sparse_tensor (experimental)
• Support for SVD and eigenvalues
• Shorthand notation dim in Tensor
• Various improvements for sparse tensors
• Support save/load on Stax nets
• Added tensor.T to transpose a tensor. This switches primal/dual dims.
• Added functions ravel_index, d2i and aliases length, rand, randn.
• Shapes can now be stacked using stack
• unpack_dim can now be used with non-uniform targets

1.7.4

NumPy 2.0 fixes, improved support for sparse tensors.

pip install phiml==1.7.4

1.7.2

Fixes NumPy compatibility and other bug fixes, adds convenience features.

1.7.1

Bug fixes, improved CompactSparseTensor support.

1.7.0

New features

• Sparse SciPy and ML tensors can now be wrapped like regular tensors.
• Shorthand shape & dual to add corresponding dual dims
• Added experimental compact sparse tensor
• Removing dims from a Shape can now be done using the - operator.
• Generic type conversion via Shape.as_type().

Improvements

• scatter() is now more flexible with treat_as_batch argument.
• Improvements to linear tracing. Improved rank deficiency detection. Linear solves will only use matrix_offset if confirmed by user.
• minimum and maximum can now be used with None values.
• Stacked trees may now include None values.
• reshaped_native() and reshaped_numpy() now support () / None for singleton dims