Skip to content

Latest commit

 

History

History

jet_substructure

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
config
config
 
 
README.md
README.md
 
 
dataset.py
dataset.py
 
 
dataset_dump.py
dataset_dump.py
 
 
models.py
models.py
 
 
neq2lut.py
neq2lut.py
 
 
requirements.txt
requirements.txt
 
 
train.py
train.py
 
 

README.md

LogicNets for Jet-Substructure Classification

This example shows the accuracy that is attainable using the LogicNets methodology on the jet substructure classification task described in our 2020 FPL paper. This example is a reimplementation of that work.

Prerequisites

  • LogicNets
  • h5py
  • yaml<6.0
  • numpy
  • pandas
  • scikit-learn

Installation

If you're using the docker image, all the above prerequisites will be already installed. Otherwise, you can install the above dependencies with pip and/or conda.

Download the Dataset

In order to download the dataset, browse to the directory where this example is contained (e.g., cd /path/to/logicnets/examples/jet_substructure/) and run the following:

mkdir -p data
wget https://cernbox.cern.ch/index.php/s/jvFd5MoWhGs1l5v/download -O data/processed-pythia82-lhc13-all-pt1-50k-r1_h022_e0175_t220_nonu_truth.z

Usage

To train the "JSC-S", "JSC-M" and "JSC-L" networks described in our 2020 FPL paper, run the following:

python train.py --arch <jsc-s|jsc-m|jsc-l> --log-dir ./<jsc_s|jsc_m|jsc_l>/

To then generate verilog from this trained model, run the following:

python neq2lut.py --arch <jsc-s|jsc-m|jsc-l> --checkpoint ./<jsc_s|jsc_m|jsc_l>/best_accuracy.pth --log-dir ./<jsc_s|jsc_m|jsc_l>/verilog/ --add-registers

Results

Your results may vary slightly, depending on your system configuration. The following results are attained when training on a CPU and synthesising with Vivado 2019.2:

Network Architecture Test Accuracy (%) LUTs Flip Flops Fmax (Mhz) Latency (Cycles)
JSC-S 69.8 244 270 1353.18 5
JSC-M 72.1 15526 881 576.70 5
JSC-L 73.1 36415 2790 389.86 6

Note, the model architectures reflect the architectures described in our FPL'20 paper.

Citation

If you find this work useful for your research, please consider citing our paper below:

@inproceedings{umuroglu2020logicnets,
  author = {Umuroglu, Yaman and Akhauri, Yash and Fraser, Nicholas J and Blott, Michaela},
  booktitle = {Proceedings of the International Conference on Field-Programmable Logic and Applications},
  title = {LogicNets: Co-Designed Neural Networks and Circuits for Extreme-Throughput Applications},
  year = {2020},
  pages = {291-297},
  publisher = {IEEE Computer Society},
  address = {Los Alamitos, CA, USA},
  month = {sep}
}