Larq-compute-engine

:tada: Features

* QuantConv2D binarized activations with tf.int32 bitpacked output (611) simonmaurer
* Automatically build macOS ARM wheels on CI (681) lgeiger

:book: Documentation

* Fix docstring of Python interpreter (677) Tombana

:construction_worker_man: Internal Improvements

* Split Python interpreter in two classes (675) Tombana

:arrow_up: Dependencies

* Bump styfle/cancel-workflow-action from 0.9.0 to 0.9.1 (678) dependabot

:tada: Features

* Extend quantiser support so as to accelerate more binary models. (668) AdamHillier
* Support for macos_arm64 platform (664) simonmaurer

:bug: Bug Fixes

* Fix default ranges in saved model converter (671) Tombana

:construction_worker_man: Internal Improvements

* Remove manylinux container from CI (670) Tombana
* Add quantization properties to python interpreter (667) Tombana
* Disable mkl for windows builds (666) AdamHillier

:tada: Features

* Add support for saved model conversion (655) lgeiger
* Add grouped binary convolution support (3/3): indirect BGEMM kernel. (551) AdamHillier
* Fuse `tfl.Quantize(lq.Dequantize(x))` -> `lq.Dequantize(x)` (644) lgeiger

:bug: Bug Fixes

* Fix windows build of grouped convolutions. (656) Tombana
* Prevent spurious int8 Quantize/Dequantize ops (646) lgeiger
* Statically link GLIBC in benchmark binaries (610) lgeiger
* Update QuantOp check (608) lgeiger

:book: Documentation

* Remove spectrum badge since we moved to GitHub discussions (616) lgeiger

:construction_worker_man: Internal Improvements

* Upload release wheels even if some jobs fail. (660) AdamHillier
* Add constant folding for `lq.quantize` and `lq.dequantize` (654) lgeiger
* Add `test-requirements.txt` file (652) lgeiger
* Add pytype linting to CI (651) lgeiger
* Fork `ConvertTFExecutorToTFLOrFlatbuffer` to re-enable argname truncation (649) lgeiger
* Add check for float32 in IsBinaryTensor (642) Tombana
* Configure the Android SDK/NDK interactively in configure.py. (627) AdamHillier
* Simplify bazel build flags on CI (626) lgeiger
* Fix lce_benchmark_model linking on Aarch32 (623) Tombana
* Add -lm link option to lce_benchmark_model (622) Tombana
* :arrow_up: CI dependencies (618) lgeiger
* Run AAR build on Ubuntu 18.04 (617) lgeiger
* Raise an error if mixed-precision models are being converted (607) CNugteren
* Automatically upload AAR package to release (603) lgeiger
* Update release.yml to actually push Windows wheels. (596) AdamHillier

:arrow_up: Dependencies

* Bump actions/cache from 2.1.5 to 2.1.6 (650) dependabot
* Pin flatbuffers to <2.0 (648) Tombana
* Upgrade TF dependency to v2.5. (628) AdamHillier
* Bump DoozyX/clang-format-lint-action from v0.11 to v0.12 (634) dependabot
* Bump actions/cache from v2.1.4 to v2.1.5 (630) dependabot
* Bump styfle/cancel-workflow-action from 0.8.0 to 0.9.0 (629) dependabot
* Bump toolmantim/release-drafter from v5.14.0 to v5.15.0 (621) dependabot
* Bump google-github-actions/setup-gcloud from v0.2.0 to v0.2.1 (612) dependabot
* Bump styfle/cancel-workflow-action from 0.7.0 to 0.8.0 (605) dependabot
* Bump toolmantim/release-drafter from v5.13.0 to v5.14.0 (606) dependabot
* Bump actions/cache from v2 to v2.1.4 (600) dependabot

:tada: Features

* Add preliminary support for the xcore target. (585) AdamHillier
* Add zero-padding support to the reference kernel (571) Tombana
* Add grouped binary convolution support (2/3): reference kernel. (550) AdamHillier
* Add grouped binary convolution support (1/3): the converter. (549) AdamHillier

:rocket: Performance

* Slightly optimise the loads in the 8x4x2 indirect BGEMM kernel (521) AdamHillier
* Add optimised indirect bgemm kernels for Aarch64. (518) AdamHillier

:bug: Bug Fixes

* Fix Windows release builds. (590) AdamHillier
* Run LCE optimize pass before and after TFLite optimize pass (525) lgeiger

:book: Documentation

* Update LCE python interpreter docstring with efficiency warning (582) jneeven

:construction_worker_man: Internal Improvements

* Fix release version number for non-manual releases. (595) AdamHillier
* Use a Python script for Bazel configuration, like TensorFlow does. (589) AdamHillier
* Add TF 2.4 to CI matrix (579) lgeiger
* Fix linebreaks in release script (577) lgeiger
* Use GitHub actions workflow dispatch to trigger nightly releases (576) lgeiger
* Fix windows build (575) lgeiger
* Only upload cache on master branch (574) lgeiger
* Do not pin official GitHub actions (572) lgeiger
* Do not pin official GitHub actions (570) lgeiger
* Set --linkopt=-dead_strip when building mac wheels (566) lgeiger
* Remove unused dependencies (565) lgeiger
* Fix filter tensor shapes in bconv2d 'death' tests. (561) AdamHillier
* Create a lightweight `BConv2DParams` struct in `/core`. (560) AdamHillier
* Fix manylinux release build (558) lgeiger
* Cancel workflows of outdated commits (556) lgeiger
* Speedup end2end tests (552) lgeiger
* :arrow_up: larq_zoov2 (547) lgeiger
* Auto-generate the TFLite Python schema. (546) AdamHillier
* Add __version__ property. (543) AdamHillier
* Use test sharding to parallelise the tests. (544) AdamHillier
* Update Flatbuffer schema for TF 2.4 (536) lgeiger
* Move padding conversion function to kernels/utils.h (534) lgeiger
* Add macros to help GCC with branch prediction (530) lgeiger
* Prebuild the benchmark binary with -O3 on Android (532) lgeiger
* Change qemu test scripts by using a new bazel rule (533) Tombana
* Upgrade TensorFlow to latest version (529) lgeiger
* Inline kernel utils functions that convert padding and activation enums (528) lgeiger
* Initialise batch norm stats for e2e models (524) AdamHillier
* Add support for batched inputs in the indirect BGEMM kernels (519) AdamHillier
* Add optimised 'Indirect BGEMM' binary convolution kernels. (516) AdamHillier
* Simplify the bconv2d padding functor. (515) AdamHillier
* Reorganise `/core` directory. (514) AdamHillier

:arrow_up: Dependencies

* :arrow_up: TensorFlow v2.4.1. (587) AdamHillier
* Bump package versions on CI. (586) AdamHillier
* Bump styfle/cancel-workflow-action from 0.6.0 to 0.7.0 (584) dependabot
* Bump toolmantim/release-drafter from v5.12.1 to v5.13.0 (580) dependabot
* :arrow_up: tensorflow2.4 (578) lgeiger
* Bump GoogleCloudPlatform/github-actions from 0.1.3 to v0.2.0 (569) dependabot
* Bump actions/upload-artifact from v2.2.0 to v2.2.1 (568) dependabot
* :arrow-up: TensorFlow submodule (564) lgeiger
* :arrow_up: tensorflow2.4.0rc1 (563) lgeiger
* Update bazelisk version on CI and in containers. (559) AdamHillier
* :arrow_up: tensorflowv2.4.0-rc0 (557) lgeiger
* :arrow_up: lint dependencies (555) lgeiger
* :arrow_up: tensorflowr2.4 (545) lgeiger
* Bump toolmantim/release-drafter from v5.12.0 to v5.12.1 (539) dependabot
* Bump toolmantim/release-drafter from v5.11.0 to v5.12.0 (538) dependabot
* Bump DoozyX/clang-format-lint-action from v0.10 to v0.11 (537) dependabot
* Bump actions/setup-python from v2.1.3 to v2.1.4 (531) dependabot
* Bump actions/upload-artifact from v2.1.4 to v2.2.0 (522) dependabot
* Bump actions/setup-python from v2.1.2 to v2.1.3 (520) dependabot
* ⬆️ tensorflow2.3.1 (517) lgeiger

:tada: Features

* Add support for multi threaded interpreter (512) lgeiger
* Add support for iterators in Interpreter.predict (511) lgeiger
* Add Python wrapper for LCE interpreter (507) lgeiger

:construction_worker_man: Internal Improvements

* Remove redundant code paths: reference BGemm, BGemm functor. (510) AdamHillier
* Make Interpreter::get_shapes and get_types more generic (509) lgeiger
* Add padding to temporary arrays to ensure we don't read beyond bounds. (505) AdamHillier
* Simplify MAKE_ZERO macro (503) lgeiger

:arrow_up: Dependencies

* Bump DoozyX/clang-format-lint-action from v0.9 to v0.10 (504) dependabot

:bug: Bug Fixes

* Fix bug introduced by 497. Make sure OPT kernels run. (501) AdamHillier