peal.data.datasets.SymbolicDataset

class peal.data.datasets.SymbolicDataset

Bases: PealDataset

Tabular (symbolic) dataset read from <dataset_path>/data.csv.

Every row becomes a float tensor in data[key]; by default the last column is the target and all others are inputs, unless task_config selects columns via x_selection/y_selection. A pandas copy of the table is kept in df (used by DiCEExplainer). The config’s tabular_preprocessing list may contain "minmax_-1_1", which scales continuous columns (more than 10 unique values, excluding targets and confounding_factors) to [-1, 1].

Parameters:

• mode (str) – Split to load: "train", "val", "test" or "all".

• config (DataConfig) – Data config; uses dataset_path, set_negative_to_zero, output_size, confounding_factors and tabular_preprocessing.

• transform (callable, optional) – Stored but not applied to the tensors. Defaults to ToTensor().

• task_config (TaskConfig, optional) – Column selection and output_channels.

• **kwargs – Ignored.

Variables:

• attributes (list of str) – Column names.

• keys (list) – Row identifiers of this split.

• df (pandas.DataFrame) – The (possibly preprocessed) table.

Methods

disable_groups	Stop returning has_confounder.
disable_hints	Clear the hints_enabled flag.
disable_idx	Stop returning the sample index.
enable_groups	Return has_confounder in dict mode (second confounding factor).
enable_hints	Set the hints_enabled flag (tabular data has no hints to return).
enable_idx	Also return the sample index from __getitem__.
generate_contrastive_collage	Render bar-chart collages of factual vs.
serialize_dataset	Write samples as a CSV next to output_dir.

Attributes

output_size

task_config.output_channels when targets are selected, else config.

__init__(mode, config, transform=torchvision.transforms.ToTensor, task_config=None, **kwargs)

Parse data.csv, build the dataframe and apply tabular preprocessing.

property output_size

task_config.output_channels when targets are selected, else config.

enable_hints()

Set the hints_enabled flag (tabular data has no hints to return).

disable_hints()

Clear the hints_enabled flag.

enable_groups()

Return has_confounder in dict mode (second confounding factor).

disable_groups()

Stop returning has_confounder.

enable_idx()

Also return the sample index from __getitem__.

disable_idx()

Stop returning the sample index.

generate_contrastive_collage(x_list, x_counterfactual_list, y_target_list, y_source_list, y_list, y_target_start_confidence_list, y_target_end_confidence_list, base_path, start_idx=0, y_original_teacher_list=None, y_counterfactual_teacher_list=None, feedback_list=None, **kwargs)

Render bar-chart collages of factual vs. counterfactual feature vectors.

For every pair a figure with three bar charts (original, counterfactual, difference) over the feature names is written to <base_path>/<zero-padded index>_collage.png.

Parameters:

• x_list (list of torch.Tensor) – Factual and counterfactual feature vectors.

• x_counterfactual_list (list of torch.Tensor) – Factual and counterfactual feature vectors.

• y_target_list (list) – Target class, predicted source class and ground truth per pair.

• y_source_list (list) – Target class, predicted source class and ground truth per pair.

• y_list (list) – Target class, predicted source class and ground truth per pair.

• y_target_start_confidence_list (list) – Target-class confidence before and after the edit.

• y_target_end_confidence_list (list) – Target-class confidence before and after the edit.

• base_path (str) – Output directory, created if missing.

• start_idx (int, optional) – Offset added to the running index in the filenames.

• y_original_teacher_list – Accepted for API compatibility; not drawn for tabular data.

• y_counterfactual_teacher_list – Accepted for API compatibility; not drawn for tabular data.

• feedback_list – Accepted for API compatibility; not drawn for tabular data.

• **kwargs – Ignored.

Returns:

(attribution_list, collage_paths) where attributions are |cf - x| tensors and paths are the written PNG files.

Return type:

tuple

serialize_dataset(output_dir, x_list, y_list, sample_names=None, hint_list=[], classifier=None)

Write samples as a CSV next to output_dir.

The file <output_dir>.csv gets one row per sample with the feature columns (task_config.x_selection or all but the last attribute) followed by the target column(s). Nothing is written for an empty x_list.

Parameters:

• output_dir (str) – Directory that is created; the CSV is named output_dir + ".csv".

• x_list (list of torch.Tensor) – Feature vectors.

• y_list (list) – Scalar targets.

• sample_names – Accepted for API compatibility; unused for tabular data.

• hint_list – Accepted for API compatibility; unused for tabular data.

• classifier – Accepted for API compatibility; unused for tabular data.