Skip to content

0.7.0 (2023-04-04)
Compare
Choose a tag to compare
@cbourjau cbourjau released this 04 Apr 08:10
· 97 commits to main since this release
0.7.0
0ca5281
This commit was created on GitHub.com and signed with GitHub’s verified signature. The key has expired.
GPG key ID: 4AEE18F83AFDEB23
Expired
Learn about vigilant mode.

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. The operator is redefined to be equivalent to constant(numpy.array(value), dtype), instead of a complex set of cases like before.
  • The spox.Type <= spox.Type (spox.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.
Assets 2
Loading