Gap

2.6 Gap Processors

Gap processors take Embeddings from two Sample Selections and compute a Domain Gap between them.

gap:
  processors:
    - name: klmvn_diag
      type: domain_gap
      columns:
        input: ["embedding"]
      distance:
        metric: klmvn_diag

    - name: mmd_linear
      type: domain_gap
      distance:
        metric: mmd_linear

    - name: fid
      type: domain_gap
      distance:
        metric: fid
      summary:
        collect_sum_outer: true

    - name: wasserstein_1d
      type: domain_gap
      distance:
        metric: wasserstein_1d
      summary:
        histogram:
          dims: 64
          bins: 32
          range: [-3.0, 3.0]

    - name: mmd_rbf
      type: domain_gap
      distance:
        metric: mmd_rbf
        kernel_params:
          gamma: 1.0
      summary:
        store_embeddings: true

    - name: mmd_poly
      type: domain_gap
      distance:
        metric: mmd_poly
        kernel_params:
          degree: 3.0
          gamma: 1.0
          coefficient0: 1.0
      summary:
        store_embeddings: true

    - name: pad
      type: domain_gap
      distance:
        metric: pad
        evaluator: mse    # or mae
      summary:
        store_embeddings: true

    - name: cmd
      type: domain_gap
      distance:
        metric: cmd
        k: 5
        feature_weights: [1.0]
      summary:
        store_embeddings: true

Distance Metric	Tunable Parameters	Summary Options
mmd_rbf	gamma	store_embeddings
mmd_poly	degree, gamma, coefficient0	store_embeddings
cmd	k, feature_weights	store_embeddings
wasserstein_1d	histogram.dims, histogram.bins, histogram.range	None
pad	evaluator (mse/mae)	store_embeddings
fid	None	collect_sum_outer
mmd_linear	None	None
klmvn_diag	None	None

See Domain Gap for per-parameter practical effects and interpretation thresholds.