qonnx version 1.0.0

What's Changed

Highlights

• Support for quantized floating point (FloatQuant) (#159)
• Operator versioning (#217) and Trunc v2 (#209)
• Remove QKeras converter and Brevitas dependency (#220). Converters should be implemented in companion packages.
• Onnxscript implementation of QuantToQCDQ (#191)
• Proper subgraph-traversal for qonnx model_wrapper transform function (#187)

Full list of changes

• Insert Identity nodes on given tensor and top-level graph inputs by @maltanar in #166
• Add option to disable scaled-int datatypes in InferDataTypes by @maltanar in #167
• Remove support for python 3.8 / 3.9 by @makoeppel in #172
• A first sample version of FloatQuant by @nghielme in #159
• Merged with float_quant, added exp bias, implemented FloatQuant.infer_node_datatype() by @ebby-s in #180
• Floating point quantization custom op and datatypes by @maltanar in #182
• Disable coverage in pytest by @maltanar in #183
• Update README.md by @maltanar in #184
• forward metadata by @jsmonson in #181
• Fix ModelWrapper method to preserve order of graph inputs and outputs by @auphelia in #186
• Store Metadata in Graph Proto rather than model Proto by @jsmonson in #188
• Fix SortGraph Removes Nodes Connected only to I/O by @jsmonson in #189
• Onnxscript implementation of QuantToQCDQ by @SpiritSeeker in #191
• Add proper subgraph-traversal for qonnx model_wrapper transform function by @jsmonson in #187
• Minor QuantToQCDQ improvements by @maltanar in #194
• Add apply_to_subgraph parameter to ModelWrapper.analysis() by @jsmonson in #193
• [DataType] vectorize checking if value is allowed for a datatype by @SpiritSeeker in #196
• Fix crash in GiveReadableTensorNames by @LinusJungemann in #199
• [FixedPt] add the fixed-point quantize transformation and unit tests by @SpiritSeeker in #198
• Revamp MultiThreshold by @maltanar in #201
• Fix failing tests and pytest-randomly by @maltanar in #205
• Update floatquant tests to fix Brevitas vs QONNX floatquant test by @nghielme in #207
• Namespace based customop registration by @tafk7 in #204
• Unpin Protobuf Version to Support for newer ONNX versions by @jsmonson in #213
• Opset version utilities by @maltanar in #215
• Op versioning for new registry system by @maltanar in #217
• Introduce v1 and v2 for Trunc by @maltanar in #209
• [trunc/quant_avg_pool] Update Trunc and QuantAveragePool to match how Brevitas Ops work by @nickfraser in #170
• Remove ApplyConfig by @maltanar in #218
• Remove QKeras export and Brevitas dependency to relax package versions by @maltanar in #220
• Add type checks to set_nodeattr by @LinusJungemann in #232
• Match set_nodeattr and get_nodeattr by @LinusJungemann in #231
• Add GiveUniqueParameterTensors to ModelWrapper.cleanup transformations by @iksnagreb in #168
• [InferDT] MAC-like node dtype output to fp32 if any fp32 operands by @iksnagreb in #230
• Try to propagate input data layout to outputs for FINN ops by @iksnagreb in #229
• Extract quantized biases for Conv/ConvTranspose by @auphelia in #224
• Add modelwrapper methods to get network global input and output tensor names by @rothej in #225
• Increase onnx version limit from <=1.16.1 to <=1.17.0 by @iksnagreb in #228
• Make gen_finn_dt_tensor consider the numpy type for INT and FIXED types by @iksnagreb in #118
• Try to fix errors in sphinx builds by @jmitrevs in #223
• Add type hints to base classes to fix Pythons automatic type hinting for FINN/FINN+ by @LinusJungemann in #226
• add CONTRIBUTING.md by @jmitrevs in #233
• Setup: compatibility with python 3.13 by @Giuseppe5 in #235
• [Transform] Introduce ComposedTransformation by @iksnagreb in #169
• [MultiThreshold] Generalize data layouts for node execution by @iksnagreb in #143

New Contributors

• @nghielme made their first contribution in #159
• @jsmonson made their first contribution in #181
• @SpiritSeeker made their first contribution in #191
• @LinusJungemann made their first contribution in #199
• @nickfraser made their first contribution in #170
• @rothej made their first contribution in #225
• @Giuseppe5 made their first contribution in #235

Full Changelog: v0.4.0...v1.0.0

qonnx version 0.4.0

Highlights:

• Channels-last conversion improvements and support for branching topologies (#148 #149 #163)
• New rounding modes in Quant (#134)
• QuantLSTM QCDQ representation (#79)
• Inference cost breakdown (#101)
• Channel pruning and quantization utilities (#71 #94)

Full list of merged PRs:
https://github.com/fastmachinelearning/qonnx/milestone/4?closed=1

qonnx version 0.3.0

Highlights:

• Initial support for converting between Q(C)DQ and Quant styles of quantization (#48 , #56 , #65 )
• Improvements to qonnx-exec for faster QONNX model execution and accuracy validation (#64 )
• Improvements to constant folding and subgraph execution (#47 )
• Improvements to inference cost measurement (#53 , #55 )

Full list of merged PRs:
https://github.com/fastmachinelearning/qonnx/milestone/3?closed=1

qonnx version 0.2.0

What's new in this release

• Support exporting QONNX from QKeras (#28 )
• Support exporting QONNX from HAWQ (#41)
• Tensor attributes in custom op wrappers (#24)
• Use knowledge of weight matrices to further optimize accumulator range calculation (#29)
• Support spatial -> channel rebalancing for certain convolutional kernels (#22)
• Numerous bugfixes and other improvements - see full list of merged PRs for this release