Time Series Generative Modeling (TSGM)

Create and evaluate synthetic time series datasets effortlessly

Get Started • Tutorials • Augmentations • Generators • Metrics • Datasets • Contributing • Citing

🧩 Get Started

TSGM is an open-source framework for synthetic time series dataset generation and evaluation.

The framework can be used for creating synthetic datasets (see 🔨 Generators ), augmenting time series data (see 🎨 Augmentations ), evaluating synthetic data with respect to consistency, privacy, downstream performance, and more (see 📈 Metrics ), using common time series datasets (TSGM provides easy access to more than 140 datasets, see 💾 Datasets ).

We provide:

• Documentation with a complete overview of the implemented methods,
• Tutorials that describe practical use-cases of the framework.

Install TSGM

TSGM now supports Keras 3 with multiple backend options. Choose one of the following installation methods:

Option 1: With TensorFlow backend (default)

pip install tsgm[tensorflow]

Option 2: With PyTorch backend

pip install tsgm[torch]

Option 3: With JAX backend

pip install tsgm[jax]

Option 4: With all backends

pip install tsgm[all]

Option 5: Basic installation (you'll need to install a backend separately)

pip install tsgm
# Then install your preferred backend:
# For TensorFlow: pip install tensorflow tensorflow-probability
# For PyTorch: pip install torch torchvision
# For JAX: pip install jax jaxlib

Backend Configuration

Set your preferred Keras backend using the environment variable:

# For TensorFlow backend
export KERAS_BACKEND=tensorflow

# For PyTorch backend  
export KERAS_BACKEND=torch

# For JAX backend
export KERAS_BACKEND=jax

M1 and M2 chips:

To install tsgm on Apple M1 and M2 chips:

# Install with TensorFlow backend
pip install tsgm[tensorflow]

# Or install with PyTorch backend
pip install tsgm[torch]

# Or install with JAX backend (excellent performance on M1/M2)
pip install tsgm[jax]

Note for PyTorch users on M1/M2 chips: Some operations may need CPU fallback on MPS devices. If you encounter MPS-related errors, set the environment variable:

export PYTORCH_ENABLE_MPS_FALLBACK=1

Note for JAX users: JAX provides excellent performance on M1/M2 chips and supports GPU acceleration. For optimal performance, consider installing JAX with Metal support:

pip install -U "jax[metal]"

Train your generative model

import tsgm

# ... Define hyperparameters ...
# dataset is a tensor of shape n_samples x seq_len x feature_dim

# Zoo contains several prebuilt architectures: we choose a conditional GAN architecture
architecture = tsgm.models.architectures.zoo["cgan_base_c4_l1"](
    seq_len=seq_len, feat_dim=feature_dim,
    latent_dim=latent_dim, output_dim=0)
discriminator, generator = architecture.discriminator, architecture.generator

# Initialize GAN object with selected discriminator and generator
gan = tsgm.models.cgan.GAN(
    discriminator=discriminator, generator=generator, latent_dim=latent_dim
)
gan.compile(
    d_optimizer=keras.optimizers.Adam(learning_rate=0.0003),
    g_optimizer=keras.optimizers.Adam(learning_rate=0.0003),
    loss_fn=keras.losses.BinaryCrossentropy(from_logits=True),
)
gan.fit(dataset, epochs=N_EPOCHS)

# Generate 100 synthetic samples
result = gan.generate(100)

⚓ Tutorials

• Introductory Tutorial Getting started with TSGM

• Tutorial Datasets in TSGM

• Tutorial Time Series Augmentations

• Tutorial Time Series Generation with VAEs

• Tutorial Conditional Time Series Generation with GANs

• Tutorial Evaluation of Synthetic Time Series Data

• Tutorial Model Selection

• Tutorial Multiple GPUs or TPU with TSGM

For more examples, see our tutorials.

🎨 Augmentations

TSGM provides a number of time series augmentations.

Augmentation	Class in TSGM	Reference
Gaussian Noise / Jittering	tsgm.augmentations.GaussianNoise	-
Slice-And-Shuffle	tsgm.augmentations.SliceAndShuffle	-
Shuffle Features	tsgm.augmentations.Shuffle	-
Magnitude Warping	tsgm.augmentations.MagnitudeWarping	Data Augmentation of Wearable Sensor Data for Parkinson’s Disease Monitoring using Convolutional Neural Networks
Window Warping	tsgm.augmentations.WindowWarping	Data Augmentation for Time Series Classification using Convolutional Neural Networks
DTW Barycentric Averaging	tsgm.augmentations.DTWBarycentricAveraging	A global averaging method for dynamic time warping, with applications to clustering.

🔨 Generators

TSGM implements several generative models for synthetic time series data.

Method	Link to docs	Type	Notes
Structural Time Series	sts.STS	Data-driven	Great for modeling time series when prior knowledge is available (e.g., trend or seasonality).
GAN	GAN	Data-driven	A generic implementation of GAN for time series generation. It can be customized with architectures for generators and discriminators.
WaveGAN	GAN	Data-driven	WaveGAN is the model for audio synthesis proposed in Adversarial Audio Synthesis. To use WaveGAN, set use_wgan=True when initializing the GAN class and use the zoo["wavegan"] architecture from the model zoo.
ConditionalGAN	ConditionalGAN	Data-driven	A generic implementation of conditional GAN. It supports scalar conditioning as well as temporal one.
BetaVAE	BetaVAE	Data-driven	A generic implementation of Beta VAE for TS. The loss function is customized to work well with multi-dimensional time series.
cBetaVAE	cBetaVAE	Data-driven	Conditional version of BetaVAE. It supports temporal a scalar condiotioning.
TimeGAN	TimeGAN	Data-driven	TSGM implementation of TimeGAN from paper
SineConstSimulator	SineConstSimulator	Simulator-based	Simulator-based synthetic signal that switches between constant and periodics functions.
Lotka Volterra	LotkaVolterraSimulator	Simulator-based	Simulator-based synthetic signal that switches between constant and periodics functions.
PdM Simulator	PdMSimulator	Simulator-based	Simulator of predictive maintenance with multiple pieces of equipment from paper

📈 Metrics

TSGM implements many metrics for synthetic time series evaluation. Check Section 3 from our paper for more detail on the evaluation of synthetic time series.

Metric	Link to docs	Type	Notes
Distance in the space of summary statistics	tsgm.metrics.DistanceMetric	Distance	Calculates a set of summary statistics in the original and synthetic data, and measures the distance between those.
Maximum Mean Discrepancy (MMD)	tsgm.metrics.MMDMetric	Distance	This metric calculated MMD between real and synthetic samples
Discriminative Score	tsgm.metrics.DiscriminativeMetric	Distance	The DiscriminativeMetric measures the discriminative performance of a model in distinguishing between synthetic and real datasets.
Demographic Parity Score	tsgm.metrics.DemographicParityMetric	Fairness	This metric assesses the difference in the distributions of a target variable among different groups in two datasets. Refer to this paper to learn more.
Predictive Parity Score	tsgm.metrics.PredictiveParityMetric	Fairness	This metric assesses the discrepancy in the predictive performance of a model among different groups in two datasets. Refer to this paper to learn more.
Privacy Membership Inference Attack Score	tsgm.metrics.PrivacyMembershipInferenceMetric	Privacy	The metric measures the possibility of membership inference attacks.
Spectral Entropy	tsgm.metrics.EntropyMetric	Diversity	Calculates the spectral entropy of a dataset or tensor as a sum of individual entropies.
Shannon Entropy	tsgm.metrics.ShannonEntropyMetric	Diversity	Shannon Entropy calculated over the labels of a dataset.
Pairwise Distance	tsgm.metrics.PairwiseDistanceMetric	Diversity	Measures pairwise distances in a set of time series.
Downstream Effectiveness	tsgm.metrics.DownstreamPerformanceMetric	Downstream Effectiveness	The downstream performance metric evaluates the performance of a model on a downstream task. It returns performance gains achieved with the addition of synthetic data.
Qualitative Evaluation	tsgm.utils.visualization	Qualitative	Various tools for visual assessment of a generated dataset.

💾 Datasets

Dataset	API	Description
UCR Dataset	tsgm.utils.UCRDataManager	https://www.cs.ucr.edu/%7Eeamonn/time_series_data_2018/
Mauna Loa	tsgm.utils.get_mauna_loa()	https://gml.noaa.gov/ccgg/trends/data.html
EEG & Eye state	tsgm.utils.get_eeg()	https://archive.ics.uci.edu/ml/datasets/EEG+Eye+State
Power consumption dataset	tsgm.utils.get_power_consumption()	https://archive.ics.uci.edu/ml/datasets/individual+household+electric+power+consumption
Stock data	tsgm.utils.get_stock_data(ticker_name)	Gets historical stock data from YFinance
COVID-19 over the US	tsgm.utils.get_covid_19()	Covid-19 distribution over the US
Energy Data (UCI)	tsgm.utils.get_energy_data()	https://archive.ics.uci.edu/ml/datasets/Appliances+energy+prediction
MNIST as time series	tsgm.utils.get_mnist_data()	https://en.wikipedia.org/wiki/MNIST_database
Samples from GPs	tsgm.utils.get_gp_samples_data()	https://en.wikipedia.org/wiki/Gaussian_process
Physionet 2012	tsgm.utils.get_physionet2012()	https://archive.physionet.org/pn3/challenge/2012/
Synchronized Brainwave Dataset	tsgm.utils.get_synchronized_brainwave_dataset()	https://www.kaggle.com/datasets/berkeley-biosense/synchronized-brainwave-dataset

TSGM provides API for convenient use of many time-series datasets (currently more than 140 datasets). The comprehensive list of the datasets in the documentation

🛠️ Contributing

We appreciate all contributions. To learn more, please check CONTRIBUTING.md.

For contributors

git clone github.com/AlexanderVNikitin/tsgm
cd tsgm
pip install -e .

Run tests:

python -m pytest

To check static typing:

mypy

💻 CLI

We provide two CLIs for convenient synthetic data generation:

• tsgm-gd generates data by a stored sample,
• tsgm-eval evaluates the generated time series.

Use tsgm-gd --help or tsgm-eval --help for documentation.

🔍 Citing

If you find this repo useful, please consider citing our paper:

@article{
  nikitin2023tsgm,
  title={TSGM: A Flexible Framework for Generative Modeling of Synthetic Time Series},
  author={Nikitin, Alexander and Iannucci, Letizia and Kaski, Samuel},
  journal={arXiv preprint arXiv:2305.11567},
  year={2023}
}

License

Apache License 2.0