at-gan 0.12.4

Training Framework for Arbitrary Tabular Generative Adversarial Networks

AT-GAN

Arbitrary Tabular Generative Adversarial Network

A Tabular GAN framework for generating synthetic tabular data from arbitrary mixed-type tabular datasets.

---

Table of Contents

1. Overview
2. Key Features
3. Installation
4. CLI Usage
5. API Usage
6. Configuration Reference
7. In-Training Evaluation Suite
8. Synthetic Data Evaluation (Post-Training)

---

Overview

at-gan is a framework for training Generative Adversarial Networks on arbitrary tabular data. It is designed to work with continuous, binary, discrete count, and categorical features within a single pipeline.

The framework combines a multi-branch generator (G), a PacGAN-style discriminator (D), an integrated evaluation suite, and Weights & Biases (W&B) sweep orchestration, experiment tracking, and training monitoring + visualization.

Goal: Training a GAN that is capable of producing realistic synthetic tabular data from a given dataset with minimal manual tuning and a transparent, observable training process.

---

Key Features

Dynamic, Config-Driven Architectures

• Generator and Discriminator built entirely from YAML-config.
• Configurable amount of layers and units.
• Configurable activations: relu, leaky_relu, elu, or any other activation supported in Keras.
• Configurable dropout layers.
• Optional Batch Normalization for G.

Mixed-Type Data Handling

• The TabularPreprocessor handles types of input features:
  • Continuous → MinMaxScaler(-1, 1) → tanh output branch.
  • Discrete Count → MinMaxScaler(0, 1) → sigmoid output branch.
  • Binary → 0/1 and optional β-distributed noise application → sigmoid output branch.
  • Categorical → One-hot encoding and optional label-preserving smoothing → softmax output branch.
• Per-column decimal precision preservation.
• Scalers and encoders are stored and reused for inference.

GAN Training and Stabilization Techniques

Technique	Controlled by	What it does
PacGAN packing	discriminator.pack_size	Concatenates k rows into a single D input → fights mode collapse
One-sided label smoothing	discriminator.label_smoothing_min	Real labels sampled from [min, 1.0] instead of hard 1.0
Label flipping	discriminator.label_flipping	Random fraction of real labels flipped to 0 to prevent D overconfidence
TTUR	g_lr / d_lr	Different LRs for G and D. Sweeps auto-clamp d_lr ≤ g_lr
G:D update ratio	g_updates_per_epoch	Multiple G steps per D step to balance the training process
LR Cosine decay + warm restarts	lr_cosine_decay	CosineDecayRestarts schedule with configurable alpha floor for the learning rate of G and D
Adam beta_1 override	adam_beta_1	Typically lowered from default 0.9 for training stability
Gradient clipping	always-on	clipnorm=1.0 on both Adam optimizers

In-Training Evaluation Suite

Runs every eval_frequency epochs on held-out real samples, logs results to W&B, and saves the best checkpoint by error score. See In-Training Evaluation Suite.

Experiment Tracking

Weights & Biases integration:

• Per-epoch loss/metric logging via a dedicated WandbCallback.
• Training visuals: correlation heatmaps + PCA overlap scatter plots.
• Local-only mode when --no-wandb is set (uses run_id="offline_run").

Sweeps & Neural Architecture Search

• W&B sweeps for Neural Architecture Search (NAS) and Hyperparameter Optimization.
• Mechanic to resume existing W&B sweeps (and single runs).

Synthetic Data Evaluation (Post-Training)

• Privacy: Distance to Closest Record (DCR)
• Statistic Fidelity: Synthetic Data Vault (SDV)
• Utility Retention: Train on Synthetic, Test on Real (TSTR)

Usage Modes

• 🖥️ CLI: train, sweep, generate, evaluate.
• 🐍 Python API (at_gan.api): train, sweep, generate, evaluate.

---

Installation

Requirements: Python 3.10 – 3.12 and dependencies listed in pyproject.toml.

Option A: Install from PyPI (recommended)

pip install at-gan

Option B: Editable install from the GitHub Repository

1. Clone the GitHub repository
2. Run the following command:

pip install -e .

Verify installation

at-gan --help
python -c "import at_gan; print(at_gan.__version__)"

Weights & Biases Login (one-time)

wandb login

💡 You can use this framework without W&B by passing --no-wandb (CLI) or enable_wandb=False (API).

---

CLI Usage

at-gan --help

train: Run or resume a single GAN training run

Flag	Short	Default	Description
--config	-c	required	Path to the YAML experiment config
--wandb / --no-wandb	-w / -nw	--wandb	Toggle W&B tracking
--export / --no-export	-e / -ne	--export	Save .keras generator file
--generate-samples	-g	1000	Auto-generate N samples post-training

Examples:

at-gan train -c configs/config.yaml -w -e -g 5000

Note: A run can be resumed via the resume_run_id config key. See Configuration Reference.

---

sweep: Run or resume a W&B sweep

Flag	Short	Description
--base-config	-c	Baseline experiment config
--sweep-config	-s	W&B sweep config (required for new sweeps)
--count	-n	Max runs this agent will execute
--sweep-id	-id	Resume an existing sweep instead of creating one

# Launch a new 50-run sweep
at-gan sweep -c configs/config.yaml -s configs/sweep_config.yaml -n 50

# Resume an existing sweep
at-gan sweep -c configs/config.yaml -id abc123 -n 20

---

generate: Generate synthetic samples from a trained generator

Flag	Short	Description
--config	-c	YAML used during the original training run
--run-id	-r	W&B run ID or "offline_run"
--samples	-n	Number of samples to generate
--output	-o	Optional override for CSV output path

at-gan generate -c configs/config.yaml -r a1b2c3 -n 10000 -o synthetic_data.csv

Note: generate always loads best_generator.keras, not the latest.

---

evaluate: Run synthetic data evaluation (post-training)

Flag	Short	Description
--real	-r	Path to the real data CSV
--synthetic	-s	Path to the synthetic data CSV
--target	-t	Discrete target column for (optional) TSTR evaluation

at-gan evaluate -c real_data.csv -r synthetic_data.csv -t target_column

Note: The TSTR evaluation is only performed if a discrete feature (i.e. binary or categorical) is specified as the target column.

---

API Usage

The Python API exposes the same primary functions as a CLI, making it easy to integrate into existing projects.

See examples/api_example.py and examples/api_example.ipynb in the GitHub Repository for a full API usage example.

Note: The train entry point also accepts a dict instead of a path to a YAML file as input.

---

Configuration Reference

Experiments are driven by two YAML files: a base config and a sweep config.

See configs/config.yaml and configs/sweep_config.yaml in the GitHub Repository for examples and recommended default values for most datasets.

Base Config Reference

# =============================================================
#  EXPERIMENT META
# =============================================================
experiment_name: "test_experiment"   # also output directory name
resume_run_id:   null                # W&B run id to resume from checkpoint (optional)
seed:            1130                # seeds Python, NumPy, TensorFlow

# =============================================================
#  DATA
# =============================================================
data:
  dataset_path: "datasets/example.csv"
  output_path:  "experiments/"          # run artifacts found in 'output_path/experiment_name/run_id/'

  # Column routing — every column the GAN should learn MUST be listed here
  continuous_cols:     ["age", "heart_rate", "glucose"]
  binary_cols:         ["male", "smoker"]
  discrete_count_cols: ["cigs_per_day"]
  categorical_cols:    ["education"]

  # Preprocessing toggles
  treat_bin_as_cat:    false    # route binary cols through OHE + softmax
  beta_noise:          true     # Apply Beta-distributed noise on binary cols
  smooth_categorical:  true     # Apply label-preserving noise on OHE groups

# =============================================================
#  MODEL
# =============================================================
model:
  latent_dim: 32             

  generator:
    units:        [64, 64]     
    dropout:      0.0
    activation:   "relu"        # relu | leaky_relu | elu | ...
    batch_norm:   true          # BatchNorm after each Dense layer
    # negative_slope: 0.2       # used only when activation == "leaky_relu"

  discriminator:
    units:               [256, 256]
    dropout:             0.2
    activation:          "leaky_relu"
    negative_slope:      0.2
    pack_size:           3      # PacGAN packing factor (1 disables packing)
    label_smoothing_min: 0.9    # e.g. real labels ~ [0.9, 1.0]
    label_flipping:      0.05   # e.g. 5% of real labels flipped to 0 each step

# =============================================================
#  TRAINING
# =============================================================
training:
  device:               "cpu"   # "cpu" or "gpu"
  epochs:               2000
  batch_size:           512
  g_updates_per_epoch:  2       # G steps per D step

  # Optimizers
  adam_beta_1:          0.5     # GAN-stable Adam beta_1
  g_lr:                 0.0002  # G Learning Rate
  d_lr:                 0.0003  # D Learning Rate

  # LR schedule
  lr_cosine_decay:                 true
  lr_cosine_decay_restart_epochs:  2000   # restart every N epochs
  g_lr_decay_alpha:                0.1    # minimum G LR fraction (floor)
  d_lr_decay_alpha:                0.1    # minimum D LR fraction (floor)

  # Evaluation & checkpointing
  checkpoint_frequency: 100     # save "latest" every N epochs
  eval_frequency:       100     # run evaluation suite every N epochs
  test_split_pct:       0.2     # percentage of data to hold out for in-training evaluation

Sweep Config Reference

# =============================================================
#  SWEEP STRATEGY & METRICS
# =============================================================
method: bayes              

metric:
  name: Eval/Total_Error     # W&B log key
  goal: minimize

early_terminate:
  type: hyperband            # Kills unpromising runs early to save compute time
  min_iter: 300              # Don't kill any run before e.g. epoch 300
  eta: 3                     # The halving rate for the Hyperband brackets

# =============================================================
# PARAMETERS
# =============================================================
parameters:

  # Sweeps choose from a fixed set of hyperparameter values
  model.latent_dim:
    values: [ 16, 32, 64, 128, 256 ]

  # -----------------------------------------------------------
  #  Generator Architecture
  # -----------------------------------------------------------
  generator.num_hidden_layers:
    values: [ 2, 3, 4 ]
  generator.base_units:
    values: [ 32, 64, 128, 256, 512 ]
  generator.max_units:
    value: 512                         
  generator.architecture_shape:
    values: [ "block", "ascending", "descending" ] 
  
  generator.dropout:
    value: 0.0                                   # e.g. Fixed to 0.0
  generator.activation:
    values: [ 'relu', 'leaky_relu' ]
  generator.batch_norm:
    values: [ true, false ]

  # -----------------------------------------------------------
  #  Discriminator Architecture
  # -----------------------------------------------------------
  discriminator.num_hidden_layers:
    values: [ 2, 3, 4 ]
  discriminator.base_units:
    values: [ 32, 64, 128, 256, 512 ]
  discriminator.max_units:
    value: 512
  discriminator.architecture_shape:
    values: [ "block", "ascending", "descending" ]
  
  discriminator.dropout:
    values: [ 0.0, 0.2, 0.3, 0.5 ]              
  discriminator.activation:
    values: [ 'relu', 'leaky_relu' ]
  discriminator.negative_slope:
    values: [ 0.1, 0.2, 0.3 ]                    
  discriminator.pack_size:
    values: [ 1, 3 ]                             
  discriminator.label_smoothing_min:
    values: [ 0.85, 0.9, 0.95, 1.0 ]             
  discriminator.label_flipping:
    values: [ 0.0, 0.05, 0.1 ]                  

  # -----------------------------------------------------------
  #  Training Loop & Optimizers
  # -----------------------------------------------------------
  training.batch_size:
    values: [ 64, 128, 256, 512 ]
  training.g_updates_per_epoch:
    values: [ 1, 2, 3 ]                          
  training.adam_beta_1:
    values: [ 0.2, 0.5, 0.7, 0.9 ] 
  
  # Learning Rates
  training.g_lr:
    distribution: log_uniform_values
    min: 0.00001                                 
    max: 0.001                                  
  training.d_lr:
    distribution: log_uniform_values
    min: 0.000005                               
    max: 0.0005                      # at-gan ensures d_lr <= g_lr

  # Cosine Decay Warm Restart Parameters
  training.lr_cosine_decay_restart_epochs:
    distribution: int_uniform
    min: 100
    max: 1000
  training.g_lr_decay_alpha:
    distribution: log_uniform_values
    min: 0.01                                    # Decay to 1% of max LR
    max: 1                                       # No decay
  training.d_lr_decay_alpha:
    distribution: log_uniform_values
    min: 0.01
    max: 1

---

In-Training Evaluation Suite

Every eval_frequency epochs, GANCallback generates synthetic samples and runs an evaluation against the held-out real samples to guide the hyperparameter sweep:

Metric	Computation
PCA Error	First Wasserstein distance between real and synthetic data across the first five PCA components
Adversarial Error	Absolute AUC deviation of a Random Forest classifier trained to distinguish real and synthetic data (|AUC - 0.5| × 2)
Total Error	sqrt((pca_error² + adv_error²) / 2.0)

Raw errors are passed through a squashing function (1 - exp(-x)) so components ∈ [0, 1].

Visual artifacts (auto-logged to W&B)

• Correlation heatmaps: real, synthetic, and absolute difference.
• PCA scatter overlay: first two principal components of real vs. synthetic.

Synthetic Data Evaluation (Post-Training)

The evaluate command runs a comprehensive benchmark suite that assesses the quality of the synthetic data generated by the GAN:

1. Privacy (DCR): Distance to Closest Record. Measures the minimum Euclidean distance (in standard deviations) between synthetic rows and real training rows. Absence of exact memorization is guaranteed if Min. DCR > 0.
2. Statistic Fidelity (SDV): Uses the Synthetic Data Vault (sdmetrics package) to generate a Quality Report, comparing 1D marginal distributions (Column Shapes) and 2D correlations (Column Pair Trends).
3. Utility Retention (TSTR): Train on Synthetic, Test on Real.
   • Splits real data into real_train (80%) and real_test (20%).
   • Trains a TRTR baseline (RandomForest, GradientBoosting, LogisticRegression) on real_train → baseline F1 on real_test.
   • Trains TSTR models on the entire synthetic set → F1 on the same real_test.
   • Reports TSTR_Mean_F1 / TRTR_Mean_F1 × 100 (F1-Score Retention in %).