Skip to content

Latest commit

 

History

History
253 lines (139 loc) · 8.91 KB

CHANGELOG.rst

File metadata and controls

253 lines (139 loc) · 8.91 KB
Raw

Change log

0.13.0 (unreleased)

Support change

  • Support for Python 3.8 has been dropped.

Bug fix

  • Value propagation of string tensors no longer raises an erroneous ValueError in some instances.

0.12.1 (2024-06-18)

Bug fix

Other changes

  • The node-naming algorithm now has constant rather than quadratic time complexity.

0.12.0 (2024-05-16)

New feature

0.11.0 (2024-04-23)

New feature

  • The opsets ai.onnx version 21 and ai.onnx.ml version 5 (released with ONNX 1.16) are now shipped with Spox.

0.10.3 (2024-03-14)

Other changes

  • Improved node creation speed by skipping the storing of the traceback
  • :class:`spox.Var` objects may now be shallow copied. Deep copies are explicitly prohibited.

Bug fix

  • Addresses node adaptation failure when referencing a non-input name from inside a subgraph by aborting opset adaptation.

0.10.2 (2023-02-08)

Other changes

  • Added packaging as an explicit dependency.

0.10.1 (2023-02-07)

Other changes

  • Spox's compatibility with older versions of onnx has been improved.

0.10.0 (2023-02-02)

New feature

  • The opsets ai.onnx version 20 and ai.onnx.ml version 4 (ONNX 1.15) are now shipped with Spox.

Other changes

  • The validation of Node attributes has been improved and more consistent exceptions are raised if needed.
  • ONNX node attributes are now computed only once and then cached so that the values are reused for validation and building the model.

0.9.3 (2023-10-23)

New feature

  • The opsets ai.onnx version 20 and ai.onnx.ml version 4 (ONNX 1.15) are now shipped with Spox.

Bug fixes

  • Address missing Value Infos when building singleton model for shape inference.
  • Fix issue where Value Propagation failure prevents model creation/inlining.

0.9.2 (2023-10-20)

Other changes

  • Fix a deprecation warning from one of Spox's dependencies.

0.9.1 (2023-10-05)

Bug fixes

  • The node-adaption no longer fails if faced with a node that has repeating inputs.
  • Forego version adaption of inlined models if no nodes are from the default domain.

0.9.0 (2023-06-12)

New feature

  • The opset ai.onnx@19 (ONNX 1.14) is now shipped with Spox.

Bug fixes

  • The constructor for ai.onnx@18::Split is no longer generated incorrectly. No extraneous attribute is generated anymore, and the num_outputs attribute is marked as required (so that Spox can infer the number of outputs).
  • Fix an issue where op.const handled numbers in the range(INT64MAX, UINT64MAX) incorrectly

Other changes

  • Inlining now no longer adds redundant Identity nodes and supports subgraphs, thanks to reimplementing the ONNX renaming routine.

0.8.1 (2023-05-xx)

Bug fixe

  • An explicit error is now raised when local subgraph arguments are leaked to an outer scope. This may happen when the subgraph callback uses side effects saving local variables, which would produce later a confusing error message.

0.8.0 (2023-05-11)

This version is intended as a release candidate for 1.0.0.

New feature

  • Spox now explicitly sets a model's IR version (currently to version 8) rather than defaulting to the latest IR version supported by the installed onnx package. Increasing the IR version is not considered a breaking change going forward. The IR version will be increase once the ecosystem has adopted it more broadly.

Bug fixes

  • Fix an issue with resolving scopes during the build process. They manifested with a KeyError in accessing variables in scope on graphs in complex dependencies between subgraphs.
  • inline now removes all symbolic dimensions from input/output shapes (i.e. N x 2 becomes ? x 2) before inferring types to avoid inconsistent interactions. This is only a visual change of the output in some cases, as they are not compared strictly in ONNX.
  • inline now explicitly does not accept model with subgraphs and local functions. Attempting to use these would usually result in invalid models. Support for them will be added in the future.

0.7.0 (2023-04-04)

This version is intended as a release candidate for 1.0.0.

New features

  • The opset ai.onnx@18 is now shipped with Spox (version 18 of the default domain, as introduced in ONNX 1.13). To avoid code duplication, unchanged implementations are 'inherited' from the previous version.

Breaking changes

  • The typing rules of the (previously partially documented) extra constructor const have changed. Its signature is now const(npt.ArrayLike, npt.DTypeLike = None) -> Var. In particular, const of a Python float no longer becomes float32, but float64 like numpy - this is a breaking change. The operator is redefined to be equivalent to constant(numpy.array(value), dtype), instead of a complex set of cases like before.
  • The Type <= Type (Type.__le__) overload is now removed.
  • Deprecated operator constructors are now no longer generated after the version their schema was deprecated. Effectively, this means ai.onnx@17::Scatter and ai.onnx@17::Upsample (available as op.scatter and op.upsample) are no longer available in spox.opset.ai.onnx.v17. They likely were not used in practice as attempting to build deprecated operators has always failed.

Bug fixes

  • The operator constructor cum_sum now also has an alias cumsum, to mirror numpy. This alias should be preferred.

New unstable features

  • spox._future.initializer(ArrayLike, DTypeLike = None) -> Var is a new function for creating variables from a constant value. The created value is constructed with numpy.array and follows the same rules. This function is opset-independent and is implemented using initializers.

0.6.2 (2023-03-29)

Pending breaking changes

  • The previously available Type <= Type (Type.__le__) overload is deprecated and will be removed in Spox 0.7.0, as it was unintentionally public.
  • Constructors for deprecated ONNX operators (currently Scatter and Upsample) now raise a warning when they are called. They will be removed entirely in 0.7.0.

Bug fixes

  • spox.inline now correctly renames unused model inputs when building. This could previously cause invalid models to be built.
  • Array attributes are now copied when they are passed to an operator. This avoids accidentally mutating them after the operator is constructed.
  • The Loop operator now has patched type inference, so that the loop-carries in its results preserve shapes if the subgraph had them inferred.

0.6.1 (2023-03-07)

Pending breaking changes

  • An undocumented extra operator constructor (const) now raises a DeprecationWarning on float, as its behaviour will change in Spox 0.7.0 to follow that of numpy.

0.6.0 (2023-02-27)

New features

  • spox.inline was added to the public interface, allowing embedding existing ONNX models in Spox.

Other changes

  • Models now have a minimum opset version of 14 for the ai.onnx domain to avoid issues with low-versioned models in ORT and other tooling.

Breaking changes

  • The operator constructor for MatMul - mat_mul - has been renamed to matmul to follow numpy naming.

0.5.0 (2023-01-20)

New features

  • The spox.build and spox.argument functions were added enabling the building of graphs through a stable interface.

Notable changes

  • The documentation formatting inside the (auto-generated) spox.opset module was greatly improved.

0.4.0 (2023-01-16)

Breaking changes

  • Removed the Var.dtype and Var.shape properties in favor of the more explicit Var.unwrap_tensor().dtype and Var.unwrap_tensor().shape ones.

Bug fixes

  • Non-ASCII characters in constant tensors are now handled correctly.
  • The Compress operator has gained an explicit type and shape inference implementation

0.3.0 (2022-12-20)

Notable changes

  • Renamed the library to Spox
  • Reduced the public API surface to a bare minimum such that downstream packages may offer a usable and stable user experience with spox-based converters. This release is intended as a release candidate. Breaking changes may still occur if necessary.