/PolyLUT

PolyLUT is the first quantized neural network training methodology that maps a neuron to a LUT while using multivariate polynomial function learning to exploit the flexibility of the FPGA soft logic.

Primary LanguagePythonApache License 2.0Apache-2.0

PolyLUT

PolyLUT is the first quantized neural network training methodology that maps a neuron to a LUT while using multivariate polynomial function learning to exploit the flexibility of the FPGA soft logic. This project is a derivative work based on LogicNets (https://github.com/Xilinx/logicnets) which is licensed under the Apache License 2.0.

Setup

Install Vivado Design Suite

Create a Conda environment

Requirements:

  • python=3.8
  • pytorch==1.4.0
  • torchvision

Install Brevitas

conda install -y packaging pyparsing
conda install -y docrep -c conda-forge
pip install --no-cache-dir git+https://github.com/Xilinx/brevitas.git@67be9b58c1c63d3923cac430ade2552d0db67ba5

Install PolyLUT package

cd PolyLUT
pip install .

Install wandb + login

pip install wandb
wandb.login()

Summary of major modifications from LogicNets

  • The novelty of our work comes through the expansion of the feature vector x at each neuron with all the monomials up to a parametric degree. When degree = 1, PolyLUT’s behavior becomes linear and it is equivalent to LogicNets. It is thus a strict generalization of LogicNets.
  • Both PolyLUT and LogicNets are capable of training on the Jet Substructure Tagging and Cybersecurity (UNSW-NB15) datasets. Additionally, PolyLUT offers compatibility with the MNIST dataset.
  • Introducing novel model architectures, PolyLUT's distinct structures are detailed in our accompanying paper.
  • PolyLUT is tailored for optimal GPU utilization.
  • To track experiments PolyLUT uses WandB insetad of TensorBoard.
  • While LogicNets enforces an a priori sparsity by utilizing a weight mask that deactivates specific weights, PolyLUT takes a different approach. It doesn't employ a weight mask but rather utilizes a feature mask (FeatureMask), which reshapes the feature vector to incorporate only fanin features per output channel. We also introduce a new mask named PolyMask, facilitating expansion with all monomials.
  • PolyLUT introduces a completely new forward function.
  • The function "calculate_truth_tables" was adapted to align with the PolyLUT neuron structure, and it was also improved for efficiency.

Citation

Should you find this work valuable, we kindly request that you consider referencing our paper as below:

@misc{andronic2023polylut,
      title={PolyLUT: Learning Piecewise Polynomials for Ultra-Low Latency FPGA LUT-based Inference}, 
      author={Marta Andronic and George A. Constantinides},
      year={2023},
      eprint={2309.02334},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}