Domain Gap (Split + Filter)

📄 Config · 📖 Docs

Filters to the zoo source only, then splits by class_name without restricting values — all four animal classes (elephant, giraffe, lion, zebra) are carried through. Wasserstein-1D compares every class pair, producing 6 pairwise distances that answer: within the zoo environment, how separable are the four species in embedding space?

Data Loader with Filter and Split

# Filter to zoo only (one controlled environment), then split by all
# class_name values. The pipeline creates four groups:
#   zoo/elephant, zoo/giraffe, zoo/lion, zoo/zebra
dataloaders:
  loaders:
    - name: zoo
      type: parquet
      path: examples/data/samples_with_images.parquet
      batch_size: 500
      filters:                  # keep only zoo-source images
        - column: source
          values: [zoo]
      split:                    # then split by class (no values = all)
        by: class_name

Feature Extraction

features:
  outputs:
    path: examples/outputs/story_domain_gap_split_filter_features.parquet
    include:
      - sample_id
      - source
      - class_name
  processors:
    - name: embedding
      type: features_embeddings
      columns:
        input: [image_bytes]
      model:
        arch: resnet18
        n_layer_feature: -2
      infer:
        batch_size: 32
        width: 64
        height: 64
        norm_mean: [0.485, 0.456, 0.406]
        norm_std: [0.229, 0.224, 0.225]

Gap Computation

# Wasserstein-1D (Earth Mover's Distance) compares every pair of groups.
# This isolates per-class distribution shift within a single source —
# e.g., how far zoo elephants are from zoo giraffes in embedding space.
gap:
  outputs:
    path: examples/outputs/story_gap_split_filter.parquet
  processors:
    - name: wasserstein_gap
      type: domain_gap
      columns:
        input: [image_bytes_embedding]
      distance:
        metric: wasserstein_1d

Run

dqm-ml process -p examples/config/scenario/domain_gap_with_split_and_filter.yaml

Config: examples/config/scenario/domain_gap_with_split_and_filter.yaml

Output

Two files are written to examples/outputs/:

• domain_gap_split_filter_features.parquet — per-image embeddings with source/class labels
• gap_split_filter.parquet — Wasserstein score between every class pair (6 pairs)

import pandas as pd

df = pd.read_parquet("examples/outputs/story_gap_split_filter.parquet")
print(df.to_markdown(index=False))

The gap output has 6 rows — one for each unordered class pair within the zoo environment:

wasserstein_1d	selection_source	selection_target	source_target
0.075	zoo_elephant	zoo_giraffe	zoo_elephant_zoo_giraffe
0.123	zoo_elephant	zoo_lion	zoo_elephant_zoo_lion
0.092	zoo_elephant	zoo_zebra	zoo_elephant_zoo_zebra
0.087	zoo_giraffe	zoo_lion	zoo_giraffe_zoo_lion
0.065	zoo_giraffe	zoo_zebra	zoo_giraffe_zoo_zebra
0.118	zoo_lion	zoo_zebra	zoo_lion_zoo_zebra

These 6 distances can be visualised as a 4 × 4 heatmap, where each cell shows the Wasserstein-1D distance between a pair of classes (the diagonal is zero). The heatmap makes it easy to spot which classes overlap in embedding space (low values) and which are well separated (high values). In this scenario, zoo_elephant → zoo_lion (0.123) and zoo_lion → zoo_zebra (0.118) show the strongest separation, while zoo_giraffe → zoo_zebra (0.065) and zoo_elephant → zoo_giraffe (0.075) are the most overlapping pairs — consistent with large herbivores sharing similar enclosure backgrounds. A Wasserstein score near 0 would indicate complete distribution overlap; values approaching 0.2 or higher would suggest near-perfect separation. None of the pairs exceed 0.15 in this dataset, confirming that the embedding space preserves some class-specific signal but the four species are not strongly separable from appearance alone within a single controlled environment.