An open source project from Data to AI Lab at MIT.

Generative adversarial training for synthesizing tabular data.

• Free software: MIT license
• Documentation: https://DAI-Lab.github.io/TGAN
• Homepage: https://github.com/DAI-Lab/TGAN

TGAN is a tabular data synthesizer. It can generate fully synthetic data from real data. Currently, TGAN can generate numerical columns and categorical columns.

TGAN has been developed and runs on Python 3.5, 3.6 and 3.7.

Also, although it is not strictly required, the usage of a virtualenv is highly recommended in order to avoid interfering with other software installed in the system where TGAN is run.

The simplest and recommended way to install TGAN is using pip:

pip install tgan

Alternatively, you can also clone the repository and install it from sources

git clone [email protected]:DAI-Lab/TGAN.git
cd TGAN
make install

For development, you can use make install-develop instead in order to install all the required dependencies for testing and code linting.

In order to be able to sample new synthetic data, TGAN first needs to be fitted to existing data.

The input data for this fitting process has to be a single table that satisfies the following rules:

• Has no missing values.
• Has columns of types int, float, str or bool.
• Each column contains data of only one type.

An example of such a tables would be:

As you can see, this table contains 4 columns: str_column, float_column, int_column and bool_column, each one being an example of the supported value types. Notice aswell that there is no missing values for any of the rows.

NOTE: It's important to have properly identifed which of the columns are numerical, which means that they represent a magnitude, and which ones are categorical, as during the preprocessing of the data, numerical and categorical columns will be processed differently.

The output of TGAN is a table of sampled data with the same columns as the input table and as many rows as requested.

TGAN includes a few datasets to use for development or demonstration purposes. These datasets come from the UCI Machine Learning repository, and have been preprocessed to be ready to use with TGAN, following the requirements specified in the Input Format section.

These datasets can be browsed and directly downloaded from the hdi-project-tgan AWS S3 Bucket

This dataset contains a single table, with information from the census, labeled with information of wheter or not the income of is greater than 50.000 $/year. It's a single csv file, containing 199522 rows and 41 columns. From these 41 columns, only 7 are identified as continuous. In TGAN this dataset is called census.

This dataset contains a single table with cartographic information labeled with the different forrest cover types. It's a single csv file, containing 465588 rows and 55 columns. From these 55 columns, 10 are identified as continuous. In TGAN this dataset is called covertype.

In this short tutorial we will guide you through a series of steps that will help you getting started with the most basic usage of TGAN in order to generate samples from a given dataset.

NOTE: The following examples are also covered in a Jupyter notebook, which you can execute by running the following commands inside your virtualenv:

pip install jupyter
jupyter notebook examples/Usage_Example.ipynb

The first step is to load the data wich we will use to fit TGAN. In order to do so, we will first import the function tgan.data.load_data and call it with the name of the dataset that we want to load.

In this case, we will load the census dataset, which we will use during the subsequent steps, and obtain two objects:

1. data, that will contain a pandas.DataFrame with the table of data from the census dataset ready to be used to fit the model.

2. continuous_columns, that will contain a list with the indices of continuous columns.

>>> from tgan.data import load_demo_data
>>> data, continuous_columns = load_demo_data('census')
>>> data.head(3).T[:10]
                              0                                     1                             2
0                            73                                    58                            18
1               Not in universe        Self-employed-not incorporated               Not in universe
2                             0                                     4                             0
3                             0                                    34                             0
4          High school graduate            Some college but no degree                    10th grade
5                             0                                     0                             0
6               Not in universe                       Not in universe                   High school
7                       Widowed                              Divorced                 Never married
8   Not in universe or children                          Construction   Not in universe or children
9               Not in universe   Precision production craft & repair               Not in universe

>>> continuous_columns
[0, 5, 16, 17, 18, 29, 38]

The next step is to import TGAN and create an instance of the model.

To do so, we need to import the tgan.model.TGANModel class and call it with the continuous_columns as unique argument.

This will create a TGAN instance with the default parameters:

>>> from tgan.model import TGANModel
>>> tgan = TGANModel(continuous_columns)

Once you have a TGAN instance, you can proceed to call it's fit method passing the data that you loaded before in order to start the fitting process:

>>> tgan.fit(data)

This process will not return anything, however, the progress of the fitting will be printed in the screen.

NOTE Depending on the performance of the system you are running, and the parameters selected for the model, this step can take up to a few hours.

After the model has been fitted, you are ready to generate new samples by calling the sample method of the TGAN instance passing it the desired amount of samples:

>>> num_samples = 1000
>>> samples = tgan.sample(num_samples)
>>> samples.head(3).T[:10]
                                         0                                     1                                   2
0                                       12                                    27                                  56

0                                       12                                    27                                  56
1                          Not in universe        Self-employed-not incorporated                             Private
2                                        0                                     4                                  35
3                                        0                                    34                                  22
4                                 Children            Some college but no degree          Some college but no degree
5                                        0                                     0                                 500
6                          Not in universe                       Not in universe                     Not in universe
7                            Never married       Married-civilian spouse present     Married-civilian spouse present
8              Not in universe or children                          Construction   Finance insurance and real estate
9                          Not in universe   Precision production craft & repair      Adm support including clerical

The returned object, samples, is a pandas.DataFrame containing a table of synthetic data with the same format as the input data and 1000 rows as we requested.

In the steps above we saw that the fitting process can take a lot of time, so we probably would like to avoid having to fit every we want to generate samples. Instead we can fit a model once, save it, and load it every time we want to sample new data.

If we have a fitted model, we can save it by calling it's save method, that only takes as argument the path where the model will be stored. Similarly, the TGANModel.load allows to load a model stored on disk by passing as argument the path where the model is stored.

>>> model_path = 'models/mymodel.pkl'
>>> tgan.save(model_path)
Model saved successfully.

Bear in mind that in case the file already exists, TGAN will avoid overwritting it unless the force=True argument is passed:

>>> tgan.save(model_path)
The indicated path already exists. Use `force=True` to overwrite.

In order to do so:

>>> tgan.save(model_path, force=True)
Model saved successfully.

Once the model is saved, it can be loaded back as a TGAN instance by using the TGANModel.load method:

>>> new_tgan = TGANModel.load(model_path)
>>> new_samples = new_tgan.sample(num_samples)
>>> new_samples.head(3).T[:10]

                                         0                                     1                                   2
0                                       12                                    27                                  56

0                                       12                                    27                                  56
1                          Not in universe        Self-employed-not incorporated                             Private
2                                        0                                     4                                  35
3                                        0                                    34                                  22
4                                 Children            Some college but no degree          Some college but no degree
5                                        0                                     0                                 500
6                          Not in universe                       Not in universe                     Not in universe
7                            Never married       Married-civilian spouse present     Married-civilian spouse present
8              Not in universe or children                          Construction   Finance insurance and real estate
9                          Not in universe   Precision production craft & repair      Adm support including clerical

At this point we could use this model instance to generate more samples.

In the previous steps we used some demonstration data but we did not show you how to load your own dataset.

In order to do so you will need to generate a pandas.DataFrame object from your dataset. If your dataset is in a csv format you can do so by using pandas.read_csv and passing to it the path to the CSV file that you want to load.

Additionally, you will need to create 0-indexed list of columns indices to be considered continuous.

For example, if we want to load a local CSV file, path/to/my.csv, that has as continuous columns their first 4 columns, that is, indices [0, 1, 2, 3], we would do it like this:

>>> import pandas as pd
>>> data = pd.read_csv('data/census.csv')
>>> continuous_columns = [0, 1, 2, 3]

Now you can use the continuous_columns to create a TGAN instance and use the data to fit it, like we did before:

>>> from tgan.model import TGANModel
>>> tgan = TGANModel(continuous_columns)
>>> tgan.fit(data)

If you want to change the default behavior of TGANModel, such as as different batch_size or num_epochs, you can do so by passing different arguments when creating the instance.

• continous_columns (list[int], required): List of columns indices to be considered continuous.
• output (str, default=output): Path to store the model and its artifacts.

• max_epoch (int, default=100): Number of epochs to use during training.
• steps_per_epoch (int, default=10000): Number of steps to run on each epoch.
• save_checkpoints(bool, default=True): Whether or not to store checkpoints of the model after each training epoch.
• restore_session(bool, default=True): Whether or not continue training from the last checkpoint.
• batch_size (int, default=200): Size of the batch to feed the model at each step.
• z_dim (int, default=100): Number of dimensions in the noise input for the generator.
• noise (float, default=0.2): Upper bound to the gaussian noise added to categorical columns.
• l2norm (float, default=0.00001): L2 reguralization coefficient when computing losses.
• learning_rate (float, default=0.001): Learning rate for the optimizer.
• num_gen_rnn (int, default=400): Number of units in rnn cell in generator.
• num_gen_feature (int, default=100): Number of units in fully connected layer in generator.
• num_dis_layers (int, default=2): Number of layers in discriminator.
• num_dis_hidden (int, default=200): Number of units per layer in discriminator.
• optimizer (str, default=AdamOptimizer): Name of the optimizer to use during fit, possible values are: [GradientDescentOptimizer, AdamOptimizer, AdadeltaOptimizer].

If you wanted to create an identical instance to the one created on step 2, but passing the arguments in a explicit way, this can be achieved with the following lines:

>>> from tgan.model import TGANModel
>>> tgan = TGANModel(
   ...:     continuous_columns,
   ...:     output='output',
   ...:     max_epoch=5,
   ...:     steps_per_epoch=10000,
   ...:     save_checkpoints=True,
   ...:     restore_session=True,
   ...:     batch_size=200,
   ...:     z_dim=200,
   ...:     noise=0.2,
   ...:     l2norm=0.00001,
   ...:     learning_rate=0.001,
   ...:     num_gen_rnn=100,
   ...:     num_gen_feature=100,
   ...:     num_dis_layers=1,
   ...:     num_dis_hidden=100,
   ...:     optimizer='AdamOptimizer'
   ...: )

We include a command-line interface that allows users to access TGAN functionality. Currently only one action is supported.

To run random searchs for the best model hyperparameters for a given dataset, we will need:

• A dataset, in a csv file, without any missing value, only columns of type bool, str, int or float and only one type for column, as specified in the Input Format.

• A JSON file containing the configuration for the search. This configuration shall contain:

  • name: Name of the experiment. A folder with this name will be created.
  • num_random_search: Number of iterations in hyper parameter search.
  • train_csv: Path to the csv file containing the dataset.
  • continuous_cols: List of column indices, starting at 0, to be considered continuous.
  • epoch: Number of epoches to train the model.
  • steps_per_epoch: Number of optimization steps in each epoch.
  • sample_rows: Number of rows to sample when evaluating the model.

You can see an example of such a json file in examples/config.json, which you can download and use as a template.

Once we have prepared everything we can launch the random hyperparameter search with this command:

tgan experiments config.json results.json

Where the first argument, config.json, is the path to your configuration JSON, and the second, results.json, is the path to store the summary of the execution.

This will run the random search, wich basically consist of the folling steps:

1. We fetch and split our data between test and train.
2. We randomly select the hyperparameters to test.
3. Then, for each hyperparameter combination, we train a TGAN model using the real training data T and generate a synthetic training dataset Tsynth.
4. We then train machine learning models on both the real and synthetic datasets.
5. We use these trained models on real test data and see how well they perform.

After the experiment has finished, the following can be found:

• A JSON file, in the example above called results.json, containing a summary of the experiments. This JSON will contain a key for each experiment name, and on it, an array of length num_random_search, with the selected parameters and its evaluation score. For a configuration like the example, the summary will look like this:

{
    'census': [
        {
            "steps_per_epoch" : 10000,
            "num_gen_feature" : 300,
            "num_dis_hidden" : 300,
            "batch_size" : 100,
            "num_gen_rnn" : 400,
            "score" : 0.937802280415988,
            "max_epoch" : 5,
            "num_dis_layers" : 4,
            "learning_rate" : 0.0002,
            "z_dim" : 100,
            "noise" : 0.2
        },
        ... # 9 more nodes
    ]
}

• A set of folders, each one names after the name specified in the JSON configuration, contained in the experiments folder. In each folder, sampled data and the models can be found. For a configuration like the example, this will look like this:

experiments/
  census/
    data/       # Sampled data with each of the models in the random search.
    model_0/
      logs/     # Training logs
      model/    # Tensorflow model checkpoints
    model_1/    # 9 more folders, one for each model in the random search
    ...

The first TAGN version was built as the supporting software for the Synthesizing Tabular Data using Generative Adversarial Networks paper by Lei Xu and Kalyan Veeramachaneni.

The exact version of software mentioned in the paper can be found in the releases section as the research pre-release

For more details about TGAN and all its possibilities and features, please check the project documentation site!

If you use TGAN for yor research, please consider citing the following paper (https://arxiv.org/pdf/1811.11264.pdf):

If you use TGAN, please cite the following work:

Lei Xu, Kalyan Veeramachaneni. 2018. Synthesizing Tabular Data using Generative Adversarial Networks.

@article{xu2018synthesizing,
  title={Synthesizing Tabular Data using Generative Adversarial Networks},
  author={Xu, Lei and Veeramachaneni, Kalyan},
  journal={arXiv preprint arXiv:1811.11264},
  year={2018}
}