Download tests/test_splits_taxonomy.py from velvet-pine-22/PULSE-code: direct link, hf CLI and curl.
- Browser
- Download file 4.95 kB
-
https://huggingface.co/velvet-pine-22/PULSE-code/resolve/main/tests/test_splits_taxonomy.py
- Command line
-
hf download hf://velvet-pine-22/PULSE-code/tests/test_splits_taxonomy.py
-
curl -L -o test_splits_taxonomy.py https://huggingface.co/velvet-pine-22/PULSE-code/resolve/main/tests/test_splits_taxonomy.py
4.95 kB
| import json | |
| from collections import Counter | |
| import pytest | |
| from conftest import REPO | |
| from pulse.split_builder import allocate, choose_inner_val, round_half_up | |
| from pulse.splits import SplitError, load_split, validate_split | |
| from pulse.taxonomy import (FINE_TO_COMPOSITE, VERB_FINE, Taxonomy, TaxonomyError, | |
| build_taxonomy) | |
| # Test folds of the source files (results/t1_cv/cv_folds.json md5 8f20127a..., | |
| # results/t2_cv/cv_folds.json md5 3b93df5d...), copied here to pin them. | |
| SOURCE_FOLDS = { | |
| "pool33": [["v1", "v12", "v19", "v24", "v29", "v34", "v39"], | |
| ["v2", "v13", "v20", "v25", "v30", "v35", "v40"], | |
| ["v3", "v15", "v21", "v26", "v31", "v36", "v41"], | |
| ["v9", "v16", "v22", "v27", "v32", "v37"], | |
| ["v11", "v17", "v23", "v28", "v33", "v38"]], | |
| "pool27": [["v25", "v28", "v11", "v38", "v5", "v35"], | |
| ["v3", "v32", "v2", "v13", "v36", "v23"], | |
| ["v26", "v33", "v22", "v41", "v24"], | |
| ["v37", "v29", "v12", "v39", "v27"], | |
| ["v1", "v40", "v31", "v34", "v30"]], | |
| } | |
| SOURCE_MD5 = {"pool33": "8f20127a120ad516bedc52c983e82eb2", | |
| "pool27": "3b93df5dcefc2f74d444cda3f5bdd997"} | |
| N_VAL = {"pool33": 4, "pool27": 3} | |
| NO_PRESSURE_33 = {"v9", "v15", "v16", "v17", "v19", "v20", "v21"} | |
| def test_committed_split(name): | |
| sp = load_split(REPO / "splits" / f"{name}.json") | |
| assert sp.name == name and sp.n_folds == 5 | |
| assert [list(sp.fold(k).test) for k in range(5)] == SOURCE_FOLDS[name] | |
| assert sp.data["source"]["md5"] == SOURCE_MD5[name] | |
| used = Counter() | |
| for k in range(5): | |
| f = sp.fold(k) | |
| assert len(f.inner_val) == N_VAL[name] | |
| assert set(f.inner_val).isdisjoint(f.test) and set(f.inner_val).isdisjoint(f.train) | |
| for v in f.inner_val: | |
| assert sp.data["volunteer_profiles"][v]["eligible"] | |
| used.update(f.inner_val) | |
| if name == "pool33": | |
| assert len(set(f.inner_val) & NO_PRESSURE_33) == 1 | |
| assert max(used.values()) == 1 # every volunteer is inner-val at most once | |
| # regenerate from the recorded profiles: identical | |
| folds = choose_inner_val(name, sp.volunteers, SOURCE_FOLDS[name], | |
| sp.data["volunteer_profiles"], sp.data["inner_val_rule"]["fraction"]) | |
| assert [f["inner_val"] for f in folds] == [list(sp.fold(k).inner_val) for k in range(5)] | |
| text = (REPO / "splits" / f"{name}.json").read_text(encoding="utf-8") | |
| assert '"data_root"' not in text and '"/' not in text # no machine paths | |
| def test_split_validation_rejects_overlap(): | |
| d = json.loads((REPO / "splits" / "pool33.json").read_text(encoding="utf-8")) | |
| d["folds"][0]["train"].append(d["folds"][0]["test"][0]) | |
| with pytest.raises(SplitError): | |
| validate_split(d) | |
| def test_rounding_and_allocation(): | |
| assert round_half_up(3.9) == 4 and round_half_up(4.05) == 4 and round_half_up(3.3) == 3 | |
| assert allocate(4, {"none": 6, "pressure": 20}) == {"none": 1, "pressure": 3} | |
| assert sum(allocate(3, {"a": 1, "b": 1, "c": 1, "d": 1}).values()) == 3 | |
| def _segs(): | |
| s = [] | |
| for i in range(60): | |
| s.append((f"v{i % 3 + 1}s1", "grasp", "cup", "right")) | |
| for i in range(55): | |
| s.append((f"v{i % 3 + 1}s2", "move", "mouse", "both")) | |
| s.append(("v1s3", "support", "cup", "left")) | |
| s.append(("v1s3", "place", "rare thing", "left")) | |
| return s | |
| def test_build_taxonomy_rules(): | |
| tax = build_taxonomy(_segs(), threshold=50, pool_volunteers={"p": ["v1", "v2"]}) | |
| assert tax["verbs"] == list(VERB_FINE) and len(VERB_FINE) == 17 | |
| assert tax["nouns"] == ["cup", "wired mouse"] # cup kept as written; alias applied | |
| assert tax["n_segments_dropped_verb"] == 1 and tax["counts"]["dropped_verbs"] == {"support": 1} | |
| assert tax["n_segments_dropped_noun"] == 1 and tax["n_segments_kept"] == 115 | |
| assert tax["chance"]["uniform"]["act_vn@5"] == pytest.approx(min(5, 17) * 2 / (17 * 2)) | |
| with pytest.raises(TaxonomyError): | |
| build_taxonomy(_segs() + [("v1s1", "juggle", "cup", "left")]) | |
| with pytest.raises(TaxonomyError): | |
| build_taxonomy(_segs() + [("v1s1", "grasp", "cup", "b")]) | |
| def test_committed_taxonomy(): | |
| tax = Taxonomy.load(REPO / "taxonomy" / "taxonomy_v1.1.json") | |
| assert tax.verbs == list(VERB_FINE) | |
| assert tax.f2c == FINE_TO_COMPOSITE and len(tax.verb_composite) == 6 | |
| assert "cup" in tax.nouns and "water cup" not in tax.nouns | |
| assert all(c >= 50 for c in tax.d["counts"]["noun"].values()) | |
| assert tax.d["n_annotated_volunteers"] == 36 | |
| lab, why = tax.classify("support", "cup", "left") | |
| assert lab is None and why == "dropped_verb:support" | |
| lab, why = tax.classify("grasp", "cup", "both") | |
| assert lab["verb_composite"] == tax.verb_composite.index("grasp-family") | |
| assert tax.classify("grasp", "food", "left")[1] == "rare_noun" | |