Skip to content

A pytorch implementation for FACE: A Normalizing Flow based Cardinality Estimator

License

Notifications You must be signed in to change notification settings

for0nething/FACE-A-Normalizing-Flow-based-Cardinality-Estimator

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

5 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

FACE

This is a pytorch implementation for the VLDB 2022 paper FACE: A Normalizing Flow based Cardinality Estimator [Citation]. Our codes are built based on nflows and torchquad.


Folder Structure

.
├── torchquadMy     # A modified pytorch implementation of adaptive importance sampling.
├── utils           # A wrapper for datasets, used to generate queries, define error metrics, etc.
├── data            # A directory for storing data. Downloaded data can be stored here.
├── train           # Codes for training normalizing flow models.
├── evaluate        # Evaluate trained models on real-world datasets for cardinality estimation.
├── environment.yml # Configuration file used to build conda environment.
└── README.md               

Quick Start

The real-world datasets can be downloaded from dataset link. We use power and BJAQ as concrete examples to illustrate how to use FACE for cardinality estimation.

  • Step 1: Build conda environment with conda env create -f environment.yml.
  • Step 2: Switch to the installed environment by conda activate testenv.
  • Step 3: Install modified torchquad by cd ./torchquadMy, and then pip install . .
  • Step 4: Download the datasets from dataset link, and then place the data into data directory.
  • Step 5: After properly setting the paths of datasets, models, etc, you can use the notebook files under train and evaluate directories to conduct experiments.

Notes:

  • Before running the codes, make sure the path variable PROJECT_PATH is set properly. This variable should be set as the path of the project root directory.
  • Current codes may be incompatible with machines that do not have GPUs.
  • For GPUs with memory less than 2080Ti (11GB), some parameters need to be set smaller, which will bring some performance loss.

License

The project is available under the MIT license.

Citation

If our work is helpful to you, please cite our paper:

@article{DBLP:journals/pvldb/WangCLL21,
  author    = {Jiayi Wang and
               Chengliang Chai and
               Jiabin Liu and
               Guoliang Li},
  title     = {{FACE:} {A} Normalizing Flow based Cardinality Estimator},
  journal   = {Proc. {VLDB} Endow.},
  volume    = {15},
  number    = {1},
  pages     = {72--84},
  year      = {2021},
  url       = {http://www.vldb.org/pvldb/vol15/p72-li.pdf},
  doi       = {10.14778/3485450.3485458},
  timestamp = {Thu, 21 Apr 2022 17:09:21 +0200},
  biburl    = {https://dblp.org/rec/journals/pvldb/WangCLL21.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}