Interfaces

1. Interfaces — What You Can Compute

DQM-ML provides 3 interfaces, each corresponding to a distinct function in the data quality methodology:

Interface	Purpose	Output
features	Compute a Feature on each Sample (e.g., image luminosity, embeddings)	One row per sample, enriched with new columns
metrics	Compute a Metric on a Sample Selection (e.g., completeness, diversity)	One row per selection (or per-column scores)
gap	Compute a Domain Gap between 2 Sample Selections (e.g., MMD, FID)	One row per pair of selections

Interface Configuration

Each interface groups optional shared configuration with a list of processors:

<interface>:      # features, metrics, or gap
  storage:        # optional — storage backend (overrides global)
    ...
  compute:        # optional — device, threads (overrides global)
    ...
  errors:         # optional — error policy (overrides global)
    ...
  outputs:        # optional — where to save results (see Outputs below)
    path: "output/..."
  processors:     # list of processor configurations
    - name: ...
      type: ...

Interface-level storage, compute, and errors override the corresponding global configuration for all processors within that interface.

Outputs

Each interface optionally defines an outputs section controlling where and how results are saved:

features:
  outputs:
    path: "output/features.parquet"          # file path
    include: ["sample_id", "class"]          # columns to include from source data (features only)
    exclude: ["meta_*"]                      # columns to exclude from source (wildcard supported)

metrics:
  outputs:
    path: "output/metrics.parquet"

gap:
  outputs:
    path: "output/gap_{}-{}.parquet"         # {source}_{target} inserted for pairwise
    pairwise: true                           # add selection_source / selection_target columns (default: true)

See Metrics for details on per-selection vs delta output formats.