fastmachinelearning/hawq-jet-tagging

hawq-jet-tagging

Installation

Clone the repository and HAWQ submodule.

git clone https://github.com/jicampos/hawq-jet-tagging.git
cd hawq-jet-tagging 
git submodule init 
git submodule update 

Set up conda environment.

conda env create -f environment.yml
conda activate hawq-env

Download LHC Jet dataset.

mkdir dataset

wget https://zenodo.org/record/3602254/files/hls4ml_LHCjet_100p_train.tar.gz -P dataset
wget https://zenodo.org/record/3602254/files/hls4ml_LHCjet_100p_val.tar.gz -P dataset

tar -zxf dataset/hls4ml_LHCjet_100p_train.tar.gz -C dataset
tar -zxf dataset/hls4ml_LHCjet_100p_val.tar.gz -C dataset

Training and Configuration

Predefined bit configurations are provided under config directory. For custom bit configurations, specify the layer name followed by the number of bits inside config files:

model:
    quant_input: # layer name
        activation_bit: 6
    dense_1:     # layer name
        weight_bit: 8
        bias_bit: 8

Specify the bit configuration file using --config for training (Default: config/config_6bit.yml)

python train_jettagger.py --config config/config_6bit.yml \
    --epochs 25 \
    --lr 0.01 \ 
    --data dataset \
    --batch-size 1024 \ 
    --save-path checkpoints

To train with Batch Normalization or L1 Regularization, use --batch-norm and --l1. For all training and logging options, use:

python train_jettagger.py --help