pe.callback package

class pe.callback.Callback

Bases: ABC

The abstract class that defines the callback for the synthetic data generation. These callbacks can be configured to be called after each PE iteration.

abstract __call__(syn_data)

This function is called after each PE iteration.

Parameters:

syn_data (pe.data.Data) – The pe.data.Data object of the synthetic data

class pe.callback.ComputeFID(priv_data, embedding, filter_criterion=None)

Bases: Callback

The callback that computes the Frechet Inception Distance (FID) between the private and synthetic data.

__call__(syn_data)

This function is called after each PE iteration that computes the FID between the private and synthetic data.

Parameters:

syn_data (pe.data.Data) – The synthetic data

Returns:

The FID between the private and synthetic data

Return type:

list[pe.metric_item.FloatMetricItem]

__init__(priv_data, embedding, filter_criterion=None)

Constructor.

Parameters:

• priv_data (pe.data.Data) – The private data

• embedding (pe.embedding.Embedding) – The embedding to compute the FID

• filter_criterion (dict, optional) – Only computes the metric based on samples satisfying the criterion. None means no filtering. Defaults to None

class pe.callback.ComputePrecisionRecall(priv_data, embedding, num_precision_neighbors=4, num_recall_neighbors=5, filter_criterion=None)

Bases: Callback

The callback that computes precision and recall metrics (https://arxiv.org/abs/1904.06991) between the private and synthetic data.

__call__(syn_data)

This function is called after each PE iteration that computes the FID between the private and synthetic data.

Parameters:

syn_data (pe.data.Data) – The synthetic data

Returns:

The FID between the private and synthetic data

Return type:

list[pe.metric_item.FloatMetricItem]

__init__(priv_data, embedding, num_precision_neighbors=4, num_recall_neighbors=5, filter_criterion=None)

Constructor.

Parameters:

• priv_data (pe.data.Data) – The private data

• embedding (pe.embedding.Embedding) – The embedding to compute the FID

• num_precision_neighbors (int, optional) – The number of neighbors to use for computing precision, defaults to 4 following https://github.com/marcojira/fld/tree/main

• num_recall_neighbors (int, optional) – The number of neighbors to use for computing recall, defaults to 5 following https://github.com/marcojira/fld/tree/main

• filter_criterion (dict, optional) – Only computes the metric based on samples satisfying the criterion. None means no filtering. Defaults to None

class pe.callback.ComputeTVD(priv_data, degree, num_bins=20, filter_criterion=None)

Bases: Callback

The callback that computes the Total Variation Distance (TVD) between the private and synthetic data.

__call__(syn_data)

This function is called after each PE iteration that computes the TVD between the private and synthetic data.

Parameters:

syn_data (pe.data.Data) – The synthetic data

Returns:

The TVD between the private and synthetic data

Return type:

list[pe.metric_item.FloatMetricItem]

__init__(priv_data, degree, num_bins=20, filter_criterion=None)

Constructor.

Parameters:

• priv_data (pe.data.Data) – The private data

• degree (int) – The degree of the TVD (e.g., 2 for 2-way TVD)

• num_bins (int, optional) – The number of bins to compute the TVD, defaults to 20

• filter_criterion (dict, optional) – Only computes the metric based on samples satisfying the criterion. None means no filtering. Defaults to None

_compute_tvd(syn_features_df, priv_features_df)

Compute the TVD between the synthetic and private features.

Parameters:

• syn_features_df (pandas.DataFrame) – The synthetic features DataFrame

• priv_features_df (pandas.DataFrame) – The private features DataFrame

Returns:

The TVD

Return type:

float

_get_features_df(data)

Get the features DataFrame from the data.

Parameters:

data (pe.data.Data) – The data

Returns:

The features DataFrame

Return type:

pandas.DataFrame

class pe.callback.ComputeWSD(priv_data, degree, num_samples=None, seed=42, filter_criterion=None)

Bases: Callback

The callback that computes the Wasserstein Distance (WSD) between the private and synthetic data.

__call__(syn_data)

This function is called after each PE iteration that computes the multiple-way WSD between the private and synthetic data.

Parameters:

syn_data (pe.data.Data) – The synthetic data

Returns:

The multiple-way WSD between the private and synthetic data

Return type:

list[pe.metric_item.FloatMetricItem]

__init__(priv_data, degree, num_samples=None, seed=42, filter_criterion=None)

Constructor.

Parameters:

• priv_data (pe.data.Data) – The private data

• degree (int) – The degree of the WSD (e.g., 2 for 2-way WSD)

• num_samples (int, optional) – The number of samples to use for the WSD for both private and synthetic data for computation efficiency. If None, all samples are used..

• seed (int, optional) – The seed to use for for sampling the data.

• filter_criterion (dict, optional) – Only computes the metric based on samples satisfying the criterion. None means no filtering. Defaults to None

_compute_wsd(syn_features_df, priv_features_df)

Compute the multiple-way WSD between the synthetic and private features.

Parameters:

• syn_features_df (pandas.DataFrame) – The synthetic features DataFrame

• priv_features_df (pandas.DataFrame) – The private features DataFrame

Returns:

The multiple-way WSD

Return type:

float

_get_features_df(data)

Get the features DataFrame from the data.

Parameters:

data (pe.data.Data) – The data

Returns:

The features DataFrame

Return type:

pandas.DataFrame

class pe.callback.DPImageBenchClassifyImages(model_name, test_data, val_data, batch_size=256, num_epochs=50, n_splits=1, lr=0.01, lr_scheduler_step_size=20, lr_scheduler_gamma=0.2, ema_rate=0.9999, **model_params)

Bases: Callback

The callback that evaluates the classification accuracy of the synthetic data following DPImageBench (https://github.com/2019ChenGong/DPImageBench).

__call__(syn_data)

This function is called after each PE iteration that computes the downstream classification metrics.

Parameters:

syn_data (pe.data.Data) – The synthetic data

Returns:

The classification accuracy metrics

Return type:

list[pe.metric_item.FloatListMetricItem]

__init__(model_name, test_data, val_data, batch_size=256, num_epochs=50, n_splits=1, lr=0.01, lr_scheduler_step_size=20, lr_scheduler_gamma=0.2, ema_rate=0.9999, **model_params)

Constructor.

Parameters:

• model_name (str) – The name of the model to use (wrn, resnet, resnext)

• test_data (pe.data.Data) – The test data

• val_data (pe.data.Data) – The validation data

• batch_size (int, optional) – The batch size, defaults to 256

• num_epochs (int, optional) – The number of training epochs, defaults to 50

• n_splits (int, optional) – The number of splits for gradient accumulation, defaults to 1

• lr (float, optional) – The learning rate, defaults to 0.01

• lr_scheduler_step_size (int, optional) – The step size for the learning rate scheduler, defaults to 20

• lr_scheduler_gamma (float, optional) – The gamma for the learning rate scheduler, defaults to 0.2

• ema_rate (float, optional) – The rate for the exponential moving average, defaults to 0.9999

_get_data_loader(data)

Getting the data loader.

Parameters:

data (pe.data.Data) – The data object

Returns:

The data loader

Return type:

torch.utils.data.DataLoader

_get_images_and_label_from_data(data)

Getting images and labels from the data.

Parameters:

data (pe.data.Data) – The data object

Returns:

The images and labels

Return type:

tuple[np.ndarray, np.ndarray]

_get_model()

Getting the model.

Raises:

ValueError – If the model name is unknown

Returns:

The model

Return type:

torch.nn.Module

evaluate(model, ema, data_loader, criterion)

Evaluating the model.

Parameters:

• model (torch.nn.Module) – The model

• ema (pe.callback.image.dpimagebench_lib.ema.ExponentialMovingAverage) – The exponential moving average object

• data_loader (torch.utils.data.DataLoader) – The data loader

• criterion (torch.nn.Module) – The criterion

Returns:

The accuracy and loss

Return type:

tuple[float, float]

class pe.callback.SampleImages(num_images_per_class=10)

Bases: Callback

The callback that samples images from the synthetic data.

__call__(syn_data)

This function is called after each PE iteration that samples images from the synthetic data.

Parameters:

syn_data (pe.data.Data) – The pe.data.Data object of the synthetic data

Returns:

A metric item with the list of sampled images

Return type:

list[pe.metric_item.ImageListMetricItem]

__init__(num_images_per_class=10)

Constructor.

Parameters:

num_images_per_class (int, optional) – number of images to sample per class, defaults to 10

class pe.callback.SaveAllImages(output_folder, path_format='{iteration:09d}/{label_id}_{label_name}/{index}.png', tqdm_enabled=True)

Bases: Callback

The callback that saves all images.

__call__(syn_data)

This function is called after each PE iteration that saves all images.

Parameters:

syn_data (pe.data.Data) – The pe.data.Data object of the synthetic data

__init__(output_folder, path_format='{iteration:09d}/{label_id}_{label_name}/{index}.png', tqdm_enabled=True)

Constructor.

Parameters:

• output_folder (str) – The output folder that will be used to save the images

• path_format (str, optional) – The format of the image paths, defaults to “{iteration:09d}/{label_id}_{label_name}/{index}.png”

• tqdm_enabled (bool, optional) – Whether to show tqdm progress bar when saving the images, defaults to True

_save_image(image, label_name, label_id, index, iteration)

A helper function that saves an image.

class pe.callback.SaveCheckpoints(output_folder, iteration_format='09d')

Bases: Callback

The callback that saves checkpoints of the synthetic data.

__call__(syn_data)

This function is called after each PE iteration that saves checkpoints of the synthetic data.

Parameters:

syn_data (pe.data.Data) – The synthetic data

__init__(output_folder, iteration_format='09d')

Constructor.

Parameters:

• output_folder (str) – The output folder that will be used to save the checkpoints

• iteration_format (str, optional) – The format of the iteration number, defaults to “09d”

_get_checkpoint_path(iteration)

Get the checkpoint path.

Parameters:

iteration (int) – The PE iteration number

Returns:

The checkpoint path

Return type:

str

class pe.callback.SaveTabToCSV(output_folder, iteration_format='09d')

Bases: Callback

The callback that saves the synthetic tabular data to a CSV file.

__call__(syn_data)

This function is called after each PE iteration that saves the synthetic tabular data to a CSV file.

Parameters:

syn_data (pe.data.Data) – The pe.data.Data object of the synthetic data

__init__(output_folder, iteration_format='09d')

Constructor.

Parameters:

• output_folder (str) – The output folder that will be used to save the CSV files

• iteration_format (str, optional) – The format of the iteration part of the CSV paths, defaults to “09d”

_get_csv_path(iteration)

Get the CSV path.

Parameters:

iteration (int) – The PE iteration number

Returns:

The CSV path

Return type:

str

class pe.callback.SaveTextToCSV(output_folder, iteration_format='09d')

Bases: Callback

The callback that saves the synthetic text to a CSV file.

__call__(syn_data)

This function is called after each PE iteration that saves the synthetic text to a CSV file.

Parameters:

syn_data (pe.data.Data) – The pe.data.Data object of the synthetic data

__init__(output_folder, iteration_format='09d')

Constructor.

Parameters:

• output_folder (str) – The output folder that will be used to save the CSV files

• iteration_format (str, optional) – The format of the iteration part of the CSV paths, defaults to “09d”

_get_csv_path(iteration)

Get the CSV path.

Parameters:

iteration (int) – The PE iteration number

Returns:

The CSV path

Return type:

str

class pe.callback.TabClassifier(test_data, model_name='xgboost', filter_criterion=None)

Bases: Callback

Evaluate tabular classification accuracy using a tabular classifier.

__call__(syn_data)

Evaluate the tabular classifier on the test set.

Parameters:

syn_data (pe.data.Data) – The synthetic training data

Returns:

Classification accuracy metrics

Return type:

list[pe.metric_item.FloatListMetricItem]

__init__(test_data, model_name='xgboost', filter_criterion=None)

Constructor.

Parameters:

• test_data (pe.data.Data) – The test data

• model_name (str, optional) – The classifier model to use, defaults to “xgboost”

• filter_criterion (dict, optional) – Only computes the metric based on samples satisfying the criterion. None means no filtering. Defaults to None

_encoding(syn_data)

Encoding categorical and numerical columns.

Parameters:

syn_data (pe.data.Data) – The synthetic training data

Returns:

The encoded synthetic training and test data

Return type:

tuple[pe.data.Data, pe.data.Data]

_get_model()

Getting the classifier model.

Subpackages

• pe.callback.common package
  • Submodules
    • pe.callback.common.compute_fid module
      • ComputeFID
        • ComputeFID.__call__()
        • ComputeFID.__init__()
    • pe.callback.common.compute_precision_recall module
      • ComputePrecisionRecall
        • ComputePrecisionRecall.__call__()
        • ComputePrecisionRecall.__init__()
    • pe.callback.common.save_checkpoints module
      • SaveCheckpoints
        • SaveCheckpoints.__call__()
        • SaveCheckpoints.__init__()
        • SaveCheckpoints._get_checkpoint_path()
• pe.callback.image package
  • Subpackages
    • pe.callback.image.dpimagebench_lib package
      • Submodules
        • pe.callback.image.dpimagebench_lib.densenet module
          • densenet()
        • pe.callback.image.dpimagebench_lib.ema module
          • ExponentialMovingAverage
            • ExponentialMovingAverage.__init__()
            • ExponentialMovingAverage.copy_to()
            • ExponentialMovingAverage.load_state_dict()
            • ExponentialMovingAverage.restore()
            • ExponentialMovingAverage.state_dict()
            • ExponentialMovingAverage.store()
            • ExponentialMovingAverage.update()
        • pe.callback.image.dpimagebench_lib.resnet module
          • resnet()
        • pe.callback.image.dpimagebench_lib.resnext module
          • resnext()
        • pe.callback.image.dpimagebench_lib.wrn module
          • BasicBlock
            • BasicBlock.forward()
          • NetworkBlock
            • NetworkBlock.forward()
          • WideResNet
            • WideResNet.forward()
  • Submodules
    • pe.callback.image.dpimagebench_classify_images module
      • DPImageBenchClassifyImages
        • DPImageBenchClassifyImages.__call__()
        • DPImageBenchClassifyImages.__init__()
        • DPImageBenchClassifyImages._get_data_loader()
        • DPImageBenchClassifyImages._get_images_and_label_from_data()
        • DPImageBenchClassifyImages._get_model()
        • DPImageBenchClassifyImages.evaluate()
    • pe.callback.image.sample_images module
      • SampleImages
        • SampleImages.__call__()
        • SampleImages.__init__()
    • pe.callback.image.save_all_images module
      • SaveAllImages
        • SaveAllImages.__call__()
        • SaveAllImages.__init__()
        • SaveAllImages._save_image()
• pe.callback.tabular package
  • Submodules
    • pe.callback.tabular.classifier module
      • TabClassifier
        • TabClassifier.__call__()
        • TabClassifier.__init__()
        • TabClassifier._encoding()
        • TabClassifier._get_model()
    • pe.callback.tabular.compute_tvd module
      • ComputeTVD
        • ComputeTVD.__call__()
        • ComputeTVD.__init__()
        • ComputeTVD._compute_tvd()
        • ComputeTVD._get_features_df()
    • pe.callback.tabular.compute_wsd module
      • ComputeWSD
        • ComputeWSD.__call__()
        • ComputeWSD.__init__()
        • ComputeWSD._compute_wsd()
        • ComputeWSD._get_features_df()
    • pe.callback.tabular.save_tab_to_csv module
      • SaveTabToCSV
        • SaveTabToCSV.__call__()
        • SaveTabToCSV.__init__()
        • SaveTabToCSV._get_csv_path()
• pe.callback.text package
  • Submodules
    • pe.callback.text.save_text_to_csv module
      • SaveTextToCSV
        • SaveTextToCSV.__call__()
        • SaveTextToCSV.__init__()
        • SaveTextToCSV._get_csv_path()

Submodules

• pe.callback.callback module
  • Callback
    • Callback.__call__()