TabDDPM: Modelling Tabular Data with Diffusion Models

This is the official code for our paper "TabDDPM: Modelling Tabular Data with Diffusion Models" (paper)

Setup the environment

1. Install conda (just to manage the env).
2. Run the following commands

   export REPO_DIR=/path/to/the/code
   cd $REPO_DIR
   
   conda create -n tddpm python=3.9.7
   conda activate tddpm
   
   pip install torch==1.10.1+cu111 -f https://download.pytorch.org/whl/torch_stable.html
   pip install -r requirements.txt
   
   # if the following commands do not succeed, update conda
   conda env config vars set PYTHONPATH=${PYTHONPATH}:${REPO_DIR}
   conda env config vars set PROJECT_DIR=${REPO_DIR}
   
   conda deactivate
   conda activate tddpm

Running the experiments

Here we describe the neccesary info for reproducing the experimental results.
Use agg_results.ipynb to print results for all dataset and all methods.

Datasets

We upload the datasets used in the paper with our train/val/test splits (link below). We do not impose additional restrictions to the original dataset licenses, the sources of the data are listed in the paper appendix.

You could load the datasets with the following commands:

conda activate tddpm
cd $PROJECT_DIR
wget "https://www.dropbox.com/s/rpckvcs3vx7j605/data.tar?dl=0" -O data.tar
tar -xvf data.tar

File structure

tab-ddpm/ -- implementation of the proposed method
tuned_models/ -- tuned hyperparameters of evaluation model (CatBoost or MLP)

All main scripts are in scripts/ folder:

• scripts/pipeline.py are used to train, sample and eval TabDDPM using a given config
• scripts/tune_ddpm.py -- tune hyperparameters of TabDDPM
• scripts/eval_[catboost|mlp|simple].py -- evaluate synthetic data using a tuned evaluation model or simple models
• scripts/eval_seeds.py -- eval using multiple sampling and multuple eval seeds
• scripts/eval_seeds_simple.py -- eval using multiple sampling and multuple eval seeds (for simple models)
• scripts/tune_evaluation_model.py -- tune hyperparameters of eval model (CatBoost or MLP)
• scripts/resample_privacy.py -- privacy calculation

Experiments folder (exp/):

• All results and synthetic data are stored in exp/[ds_name]/[exp_name]/ folder
• exp/[ds_name]/config.toml is a base config for tuning TabDDPM
• exp/[ds_name]/eval_[catboost|mlp].json stores results of evaluation (scripts/eval_seeds.py)

To understand the structure of config.toml file, read CONFIG_DESCRIPTION.md.

Baselines:

• smote/
• CTGAN/ -- TVAE official repo
• CTAB-GAN/ -- official repo
• CTAB-GAN-Plus/ -- official repo

Examples

Run TabDDPM tuning.

Template and example (--eval_seeds is optional):

python scripts/tune_ddpm.py [ds_name] [train_size] synthetic [catboost|mlp] [exp_name] --eval_seeds
python scripts/tune_ddpm.py churn2 6500 synthetic catboost ddpm_tune --eval_seeds

Run TabDDPM pipeline.

Template and example (--train, --sample, --eval are optional):

python scripts/pipeline.py --config [path_to_your_config] --train --sample --eval
python scripts/pipeline.py --config exp/churn2/ddpm_cb_best/config.toml --train --sample

It takes approximately 7min to run the script above (NVIDIA GeForce RTX 2080 Ti).

Run evaluation over seeds
Before running evaluation, you have to train the model with the given hyperparameters (the example above).

Template and example:

python scripts/eval_seeds.py --config [path_to_your_config] [n_eval_seeds] [ddpm|smote|ctabgan|ctabgan-plus|tvae] synthetic [catboost|mlp] [n_sample_seeds]
python scripts/eval_seeds.py --config exp/churn2/ddpm_cb_best/config.toml 10 ddpm synthetic catboost 5