Brevitas

Highlights

- Fix an issue with LOG_FP or POWER_OF_TWO restrictions on the scale factors, where the absolute value of the scale was incorrectly being computed before exponentiation. This is a breaking change that restores the (correct) behaviour implemented in earlier releases of Brevitas, and it restores the full accuracy of some pretrained models like QuartzNet. Training with those settings should now be more stable too.
- Supporting solving enum directives in quantizers that default to None. This restores support for inline enum driven bias quantization that was previously silently failing.
- Add support for exporting a whole Brevitas model to ONNX (down to all its low level operations) through a call to torch.onnx.export.
- Initial support for a 'generic' proxy-level ONNX export flow, found under brevitas.export.onnx.generic.
- Initial support for exporting to Xilinx's XIR format, found under brevitas.export.onnx.vitis_ai.xir.
- Experimental novel quantized L2,inf weight norm quantizers with learned scale (Int4WeightPerTensorFloatDecoupled and Int4WeightPerTensorFixedPointDecoupled), for high accuracy with per-tensor scaling at low precision, especially on depthwise separable layers.
- Add the binary and ternary quantizers used in the bnn_pynq examples as standalone quantizers under brevitas.quant.

Changelog:
- Add support for \_\_torch_function\_\_ to QuantTensor for supported versions of PyTorch. This finally removes the main barrier in making usage of QuantTensor more ingrained into the overall library.
- Correctly export operators that are invoked through \_\_torch_function\_\_ and that are invariant to quantization such as torch.nn.functional.max_pool2d in both standard ONNX and PyXIR.

Changelog:

- Important bugfix affecting ollection of activation statistics when retraining with BREVITAS_IGNORE_MISSING_KEYS=1. Statistics where not being collected and instead it was using the default baseline value of 1.0 to initialize the scale factor. The problem doesn't affect using load_state_dict(strict=False), which is an alternative to the flag above.
- Refactor proxies and mixins and simplify a bit the assumptions under which an injector proxy can be created (i.e. always within a quantized layer).
- Release tutorial on quantizers.

Release version 0.3.0.

Changelog:
- Enum and shape solvers are now implemented through extended dependency injectors. This finally makes declarative quantizers self-contained.
- Reorganize CI.
- Various smaller features and fixes.

Release version 0.2.1.

Changelog:
- Fix a few issues when using QuantTensors w/ zero point.
- Fix Hadamard layer, the implementation had fallen behind w.r.t QuantLayer and QuantTensor semantics.
- Make sure that the training flag in a QuantTensor is always set by the Module generating it.

First release on PyPI, version 0.2.0.

bnn_pynq-r1
Updated *FC networks from maltanar with TensorNorm instead of BatchNorm as last year, to ease deployment to FINN.

quant_mobilenet_v1_4b-r2
Update pretrained MobileNet V1 w/ 4b weights in the first layer.

cnv_test_ref-r0
Reference tests vectors for CNV models, r0.

bnn_pynq-r0
CNV, LFC, SFC, TFC topologies, originally designed for BNN-PYNQ, trained with Brevitas. Thanks to maltanar and ussamazahid96 .
Matching txt files contain batch-by-batch accuracy results, taken directly from the evaluation scripts.

quant_quartznet_4b-r0
Pretrained 4b QuartzNet for automatic speech recognition.

quant_quartznet_8b-r0
Pretrained 8b QuartzNet encoder and decoder for automatic speech recognition.

quant_melgan_8b-r0
Pretrained quantized 8b MelGAN vocoder on LJSpeech.

quant_proxylessnas_mobile14_hadamard_4b-r0
Pretrained quantized ProxylessNAS Mobile14 with everything at 4b (except input and weights of the first layer at 8 bits) and an Hadamard classifier as the last layer.

quant_proxylessnas_mobile14_4b-r0
Pretrained quantized ProxylessNAS Mobile14 with everything at 4b (except input and weights of the first layer at 8 bits).

quant_proxylessnas_mobile14_4b5b-r0
Pretrained quantized ProxylessNAS Mobile14 with 5b inputs and weights in depthwise layers, and everything else at 4b (except the weights of the first layer at 8 bits).

quant_mobilenet_v1_4b-r1
Re-release pretrained quantized 4b MobileNet V1 with proper naming so that it can be downloaded automatically.


examples-0.0.1
Add pretrained .pth for quantized 4b MobileNet V1.