How jacobyhsi/tabrep works
TabRep is a research toolkit for training and comparing tabular synthetic-data generators. The principal workflow is to prepare a dataset, select a method and train or sample through the command-line entry point, then evaluate generated data. The repository also includes documented baseline generators and a low-dimensional TabSYN-style VAE/diffusion experiment. Training and sampling internals are only partly sampled, so method-to-internal-model links are shown only where the source index or excerpts support them; documented capabilities are retained without assuming their internal call graphs.
Main entry points: main.py Experiment CLI · tabrep_ddpm/train.py DDPM training · sample.py DDPM sampling · tabrep_flow/train.py Flow training
Components 30
Data and CLI
- Experiment CLI
main.py - Dataset download
download_dataset.py - Dataset processing
process_dataset.py - Shared preprocessing
utils_train.py - Dataset transforms
src/data.py - Dataset environment
src/env.py - Task metrics
src/metrics.py
TabRep Models
- DDPM training
tabrep_ddpm/train.py - Mixed diffusion
tabrep_ddpm/models/gaussian_multinomial_distribution.py - DDPM networks
tabrep_ddpm/models/modules.py - DDPM sampling
tabrep_ddpm/sample.py - Flow training
tabrep_flow/train.py - Flow matching
tabrep_flow/models/flow_matching.py - Flow networks
tabrep_flow/models/modules.py - Flow sampling
tabrep_flow/sample.py
Baseline Models
- CoDi generator
baselines/codi/diffusion_discrete.py - CTGAN and TVAE
baselines/ctgan/models - Goggle generator
baselines/goggle/GoggleModel.py - GReaT generator
baselines/great/models/great.py - SMOTE baseline
baselines/smote/main.py - STaSy generator
baselines/stasy/losses.py - TabDDPM baseline
baselines/tabddpm/models/gaussian_multinomial_distribution.py - TabSYN baseline
baselines/tabsyn/latent_utils.py - Low-dim VAE
ablations/tabsyn_lowdim/vae/model.py
Evaluation
- ML efficiency
eval/eval_mle.py - Density evaluation
eval/eval_density.py - Quality evaluation
eval/eval_quality.py - Detection evaluation
eval/eval_detection.py - Privacy evaluation
eval/eval_privacy.py
Other components
- Researcher
Connections 31
- Researcher → Dataset download: obtains data
- Researcher → Dataset processing: prepares data
- Researcher → Experiment CLI: selects method
- Dataset download → Dataset processing: provides datasets
- Experiment CLI → DDPM training: dispatches train
- Experiment CLI → DDPM sampling: dispatches sample
- Experiment CLI → Flow training: dispatches train
- Experiment CLI → Flow sampling: dispatches sample
- Experiment CLI → CoDi generator: selects baseline
- Experiment CLI → CTGAN and TVAE: selects baseline
- Experiment CLI → Goggle generator: selects baseline
- Experiment CLI → GReaT generator: selects baseline
- Experiment CLI → SMOTE baseline: selects baseline
- Experiment CLI → STaSy generator: selects baseline
- Experiment CLI → TabDDPM baseline: selects baseline
- Experiment CLI → TabSYN baseline: selects baseline
- Shared preprocessing → Dataset transforms: transforms dataset
utils_train.py - DDPM training → Shared preprocessing: preprocesses data
utils_train.py - DDPM training → DDPM networks: uses networks
tabrep_ddpm/train.py - DDPM training → Mixed diffusion: trains diffusion
tabrep_ddpm/train.py - DDPM sampling → DDPM networks: uses networks
- DDPM sampling → Mixed diffusion: samples diffusion
- Flow training → Flow matching: trains flow
- Flow training → Flow networks: uses networks
- Flow sampling → Flow matching: samples flow
tabrep_flow/sample.py - Flow sampling → Flow networks: uses networks
tabrep_flow/sample.py - Researcher → ML efficiency: evaluates output
- Researcher → Density evaluation: evaluates output
- Researcher → Quality evaluation: evaluates output
- Researcher → Detection evaluation: evaluates output
- Researcher → Privacy evaluation: evaluates output