1.5.0

The Φ-ML paper has been published! DOI

This version updates the value/variable attribute system and is compatible with TensorFlow 2.16 and the latest JAX versions.

Highlights

• Value attributes are now used in optimization and gradient computation.
• tensorflow_probability is no longer required for median() and quantile()
• Added find_differences to compare trees
• Added sort
• grid_sample and closest_grid_values now support specifying grid dims via coordinates

Notable fixes

• PyTorch's long-standing JIT adjoint linear solve problem has been resolved
• Tracing JIT functions within a gradient function no longer stores the gradients (caused JAX escaped tracers)

pip install phiml==1.5.0

1.4.0

Version 1.4 adds a lot of functionality to Φ-ML and includes a ton of fixes!

This version is archived at figshare, DOI: DOI.

Highlights

• Users now have more control over sparse matrices with to_format for sparse/dense conversion and is_sparse.
• Added find_closest() with dense and k-d tree implementations
• New functions argmin, argmax, equal, erf, incomplete_gamma
• Improved gradient descent optimizer via math.minimize()
• Support ... for remaining dimensions in reshaped_native, reshaped_tensor.
• scatter now supports modes min, max, prod, any, all
• min, max, nonzero now support sparse tensors
• random_uniform now supports setting limits by Tensor
• cross_product now retains item names
• slice now supports primitives
• map now supports dims='object' to map over layout dimensions
• Non-uniform tensors can now be used with dot, median
• vec now allows empty vectors
• linspace now supports multiple dimensions
• Sparse-sparse matrix multiplication

pip install phiml==1.4.0

1.3.1

Bug fixes, improved compatibility for all_available, gather, map

1.3.0

Highlights

• New functions at_min, at_max, neighbor_min, etc.
• New JIT-specific functions when_available and perf_counter().
• @broadcast now supports specifying keyword arguments
• Support for custom types has been improved, e.g. in gather, flatten
• iterate now allows None as initial value and skips it when stacking
• Experimental support for higher-order dual dims (internal at the moment).
• Experimental tracing of linear functions involving gather / scatter operations and sparse tensors.
• Lots of bug fixes

Breaking changes

• stack and map no longer wrap non-shapable function outputs

pip install phiml==1.3.0

1.2.1

Bug fixes.

pip install phiml==1.2.1

1.2.0

Highlights

• Add conditional extrapolation
• boolean_mask() for non-Tensor arguments and sparse matrices, Support boolean_mask() via Tensor[bool-Tensor]
• Certain operations like stack now support str instead of Shape
• Sparse matrix improvements

pip install phiml==1.2.0

1.1.0

Improvements and new features

• Improved map() and broadcast with correct kwargs handling and user-defined slicing
• Support for rotation matrices and 3D rotations
• Improved degree-radians conversions
• Extrapolation improvements, added extrapolation.get_normal() and get_tangential()
• Added Shape.as<type>(), assert_all_sizes_defined() convenience methods

pip install phiml==1.1.0

1.0.4

Minor additions and fixes.

• rotate_vector() now accepts rotation matrices
• Added extrapolation.get_normal() and get_tangential()
• Added Shape.as<type>() convenience methods

pip install phiml==1.0.4

1.0.3

Extrapolation bug fixes.

pip install phiml==1.0.3

1.0.2

Extrapolation bug fixes. Tensors can now be padded with non-constant edge values.

pip install phiml==1.0.2