Introduction

NeuDep is a tool to detect binary memory dependencies.

Installation

First, create the conda environment,

conda create -n neudep python=3.9 scipy scikit-learn requests

and activate the conda environment:

conda activate neudep

Then, install the latest PyTorch (assume you have GPU):

conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch

Enter the neudep root directory: e.g., path/to/neudep, and install neudep:

pip install --editable .

Install PyArrow for large datasets:

pip install pyarrow

For faster training install NVIDIA's apex library:

git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" \
  --global-option="--deprecated_fused_adam" --global-option="--xentropy" \
  --global-option="--fast_multihead_attn" ./

File Structure

checkpoints - store the pretrained and finetuned checkpoints. We provide the pretrained model here:

command - scripts to pretrain and finetune the model. Hyperparameters are included in the training scripts too, e.g., command/finetune/table3/finetune_table3_all.sh we have MAX_SENTENCES=32 indicating batch size is 32.

fairseq - implementation of model architecture, preprocessing pipeline, and training task loss

data-src - preprocessed dataset for model to binarize. We put the data for pretraining at here, and finetuning at here

data-bin - stores binarized dataset for actual training