Hessian AWare Quantization V3: Dyadic Neural Network Quantization

This is a brief description of HAWQV3, which is a Hessian AWare Quantization Framework, pre-recorded for the TVM Conference. HAWQV3 uses integer-only quantization with Dyadic scaling. As a result, the Neural Network inference would only include Integer multiplication, addition, and bit shifting that can be performed very efficiently. Please see below for more details on the HAWQ framework:

[HAWQV3]: https://arxiv.org/abs/2011.10680
[HAWQV2]: https://arxiv.org/abs/1911.03852
[HAWQ]: https://arxiv.org/abs/1905.03696

[Code]: https://github.com/zhen-dong/hawq

Видео Hessian AWare Quantization V3: Dyadic Neural Network Quantization канала Amir Gholaminejad