v1.1.0 — Broader model coverage

Broader model coverage and correctness fixes for onnx2coreml, the ONNX → Apple Core ML converter (.mlpackage ML Program and .mlmodel NeuralNetwork), built on coremltools' MIL builder with numerical-parity verification against ONNX Runtime.

This release brings supported operator coverage to 106 ops and improves conversion fidelity for several real-world model patterns.

Added

• SpaceToDepth lowering.
• Flat-directory batch conversion and parity-check tool.
• Project citation metadata and Core ML model provenance metadata.

Fixed

• Pow with scalar constant exponents on the NeuralNetwork backend.
• BatchNormalization for rank-2 (N, C) inputs.
• Resize linear + asymmetric sampling.
• Coverage reports now run after cleanup and fusion passes, eliminating false unsupported-op hits.

Install

uv pip install onnx2coreml==1.1.0

Full changelog: https://github.com/devin-lai/onnx2coreml/blob/main/CHANGELOG.md
Compare: v1.0.0...v1.1.0

v1.0.0 — First stable release

First stable release of onnx2coreml — an ONNX → Apple Core ML converter producing .mlpackage (ML Program / MIL) and .mlmodel (NeuralNetwork), built on coremltools' MIL builder and verified for numerical parity against ONNX Runtime.

Highlights

• 105 operator lowerings across elementwise/math, activations, convolution/pooling, normalization, linear, shape/movement, indexing/gather-scatter, reduction, recurrent, and attention families.
• Two output formats — ML Program (.mlpackage, fp16, modern on-device performance) and NeuralNetwork (.mlmodel, older-OS reach).
• Verified, not assumed — every operator has parity tests run against the live Core ML runtime on macOS.

Added

• Operator coverage expanded to 105 ops: LSTM (recurrent), GridSample, DepthToSpace, TopK, the ReduceL1/L2/LogSum/LogSumExp/SumSquare family, GatherND/GatherElements, ScatterND/ScatterElements, NonZero, matrix Inverse (com.microsoft), and the trigonometric, logical, and comparison operators.
• Selective fp16 precision: convert(..., fp32_op_types={...}) keeps numerically sensitive ops in fp32 while the rest runs fp16.
• Model-zoo evaluation and export tooling under tools/.

Fixed

• PRelu now handles slopes that broadcast across more than the channel dimension (e.g. (C, H, W)), using the exact elementwise identity rather than Core ML's per-channel prelu op.
• ConvTranspose now honors output_padding.
• PRelu handles non-rank-4 inputs and scalar slopes; Clip handles integer tensors; Expand handles boolean tensors.
• Constants are saturated into the int32 and fp16 ranges to avoid overflow (int64 sentinels and FLT_MAX attention-mask fills).

Install

uv pip install onnx2coreml==1.0.0

Full changelog: https://github.com/devin-lai/onnx2coreml/blob/main/CHANGELOG.md