GenTab

Synthetic Tabular Data Generation Library

Overview

This Python library specializes in the generation of synthetic tabular data. It has a diverse range of statistical, Machine Learning (ML) and Deep Learning (DL) methods to accurately capture patterns in real datasets and replicate them in a synthetic context. It has multiple applications including pre-processing of tabular datasets, data balancing, resampling...

Features

:nut_and_bolt: Pre-process your data.

:clock130: State-of-the-art models.

:recycle: Easy to use and customize.

Install

The gentab library is available using pip. We recommend using a virtual environment to avoid conflicts with other software on your machine.

pip install gentab

Available Generators

Below is the list of the generators currently available in the library.

Linear

Model	Example	Paper
Random Over-Sampling		link
SMOTE		link
ADASYN		link

PDF

Model	Example	Paper
Gaussian Copula		link

AE

Model	Example	Paper
TVAE		link

GAN

Model	Example	Paper
CTGAN		link
CTAB-GAN		link
CTAB-GAN+		link

Diffusion

Model	Example	Paper
ForestDiffusion		link

LLM

Model	Example	Paper
GReaT		link
Tabula		link

Hybrid

Model	Example	Papers
Copula GAN		link link
AutoDiffusion		link

Examples

Generation

from gentab.generators import AutoDiffusion
from gentab.evaluators import MLP
from gentab.data import Config, Dataset
from gentab.utils import console

config = Config("configs/playnet.json")

dataset = Dataset(config)
dataset.reduce_size(
    {
        "left_attack": 0.97,
        "right_attack": 0.97,
        "right_transition": 0.9,
        "left_transition": 0.9,
        "time_out": 0.8,
        "left_penal": 0.5,
        "right_penal": 0.5,
    }
)
dataset.merge_classes(
    {
        "attack": ["left_attack", "right_attack"],
        "transition": ["left_transition", "right_transition"],
        "penalty": ["left_penal", "right_penal"],
    }
)
dataset.reduce_mem()

console.print(dataset.class_counts(), dataset.row_count())
generator = AutoDiffusion(dataset)
generator.generate()
console.print(dataset.generated_class_counts(), dataset.generated_row_count())

evaluator = MLP(generator)
evaluator.evaluate()

dataset.save_to_disk(generator)

Tuning

from gentab.generators import AutoDiffusion
from gentab.evaluators import LightGBM
from gentab.tuners import AutoDiffusionTuner
from gentab.data import Config, Dataset

config = Config("configs/adult.json")

dataset = Dataset(config)
dataset.merge_classes({
    "<=50K": ["<=50K."], ">50K": [">50K."]
})
dataset.reduce_mem()

generator = AutoDiffusion(dataset)

evaluator = LightGBM(generator)

trials = 10
time = 60 * 60 * 8
tuner = AutoDiffusionTuner(evaluator, trials, timeout=time)
tuner.tune()
tuner.save_to_disk()

Loading Stored Synthetic Datasets

from gentab.generators import AutoDiffusion
from gentab.evaluators import LightGBM
from gentab.tuners import AutoDiffusionTuner
from gentab.data import Config, Dataset

config = Config("configs/adult.json")

dataset = Dataset(config)
dataset.merge_classes({
    "<=50K": ["<=50K."], ">50K": [">50K."]
})
dataset.reduce_mem()

# Load previously saved dataset...
generator = AutoDiffusion(dataset)
generator.load_from_disk()

# Do work with previously generated but not tuned dataset...
evaluator = LightGBM(generator)
evaluator.evaluate()
evaluator.evaluate_baseline()

# Load previously tuned and saved dataset...
tuner = AutoDiffusionTuner(evaluator, 0)
tuner.load_from_disk()

# Do work with previously tuned dataset...
evaluator.evaluate()
evaluator.evaluate_baseline()