Transforms

The cesnet-datazoo package supports configurable transforms of input data in a similar fashion to what torchvision is doing for the computer vision field. Input features are split into three groups, each having its own transformation. Those groups are PPI sequences, flow statistics, and packet histograms.

• Transformation configured in ppi_transform of DatasetConfig is applied to PPI sequences.
• flowstats_transform is applied to flow statistics (excluding boolean features, such as flow end reasons or TCP flags).
• flowstats_phist_transform is applied to packet histograms.

Transforms are implemented in a separate package CESNET Models. See cesnet_models.transforms documentation for details.

Limitations

The current implementation does not support the composing of transformations.

Available transformations

PPI sequences

• ClipAndScalePPI

Flow statistics

• ClipAndScaleFlowstats

Packet histograms

• NormalizeHistograms

More transformations will be implemented in future versions.

Data scaling

Transformations implementing data scaling will be fitted, if needed, on a subset of training data during dataset initialization.