from __future__ import annotations import json import numpy as np from psi.data.seed.omni_oss_sampler import ( _video_manifest_signature, ensure_video_split_inverse_index, ) from psi.data.seed.source_split import ( VIDEO_MANIFEST_IDENTITY_FORMAT, load_video_manifest_identity, motion_source_id, select_source_row_ids, source_leaf_key, video_manifest_identity_path, validation_source_ids, ) def test_motion_source_id_matches_across_modalities(): text = motion_source_id("seed", "g1/csv/240918/body_check_001__A548.csv") command = motion_source_id("seed", "body_check_001__A548") assert text == command == "seed/body_check_001__A548" assert source_leaf_key(r"converted_motion\00001\dance.foo") == "dance.foo" def test_source_rows_are_atomic_complementary_and_manifest_ordered(): sources = ["family/a", "family/a", "family/b", "family/c", "family/c", "family/d"] train = select_source_row_ids(sources, split="train", val_ratio=0.5, split_seed=42) val = select_source_row_ids(sources, split="val", val_ratio=0.5, split_seed=42) assert np.all(train[:-1] < train[1:]) assert np.all(val[:-1] < val[1:]) assert set(train).isdisjoint(set(val)) assert set(train) | set(val) == set(range(len(sources))) for source in set(sources): memberships = {int(row_id in set(val)) for row_id, value in enumerate(sources) if value == source} assert len(memberships) == 1 def test_source_assignment_is_independent_of_input_order(): sources = [f"source/{index:03d}" for index in range(100)] selected = validation_source_ids(sources, val_ratio=0.1, split_seed=7) reversed_selected = validation_source_ids(reversed(sources), val_ratio=0.1, split_seed=7) assert selected == reversed_selected def test_tiny_dataset_keeps_train_and_val_nonempty(): sources = ["source/a", "source/b"] val = validation_source_ids(sources, val_ratio=1e-12, split_seed=1) assert len(val) == 1 val = validation_source_ids(sources, val_ratio=1.0 - 1e-12, split_seed=1) assert len(val) == 1 def test_manifest_identity_survives_copy_and_mtime_change(tmp_path): payload = b'{"video_path":"demo.chunk00.f000000.mp4"}\n' identity = { "format": VIDEO_MANIFEST_IDENTITY_FORMAT, "bytes": len(payload), "rows": 1, "sha256": "a" * 64, } manifests = [] for index in range(2): root = tmp_path / f"mount-{index}" root.mkdir() manifest = root / "metadata.jsonl" manifest.write_bytes(payload) video_manifest_identity_path(manifest).write_text( json.dumps(identity), encoding="utf-8", ) manifests.append(manifest) manifests[1].touch() assert load_video_manifest_identity(manifests[0]) == identity assert load_video_manifest_identity(manifests[1]) == identity assert _video_manifest_signature(manifests[0]) == _video_manifest_signature(manifests[1]) def test_manifest_identity_keeps_explicit_excluded_video_sources(tmp_path): payload = b'{"video_path":"kept.chunk00.f000000.mp4"}\n' manifest = tmp_path / "metadata.jsonl" manifest.write_bytes(payload) video_manifest_identity_path(manifest).write_text( json.dumps( { "format": VIDEO_MANIFEST_IDENTITY_FORMAT, "bytes": len(payload), "rows": 1, "sha256": "c" * 64, "excluded_source_bases": ["missing-b", "missing-a"], } ), encoding="utf-8", ) identity = load_video_manifest_identity(manifest) assert identity is not None assert identity["excluded_source_bases"] == ["missing-a", "missing-b"] def test_manifest_identity_rejects_truncated_copy(tmp_path): manifest = tmp_path / "metadata.jsonl" manifest.write_bytes(b"{}\n") video_manifest_identity_path(manifest).write_text( json.dumps( { "format": VIDEO_MANIFEST_IDENTITY_FORMAT, "bytes": 999, "rows": 1, "sha256": "b" * 64, } ), encoding="utf-8", ) with np.testing.assert_raises_regex(RuntimeError, "size mismatch"): load_video_manifest_identity(manifest) def test_video_source_inverse_has_no_gaps_overlap_or_partial_sources(tmp_path): prefixes = { "a": (0, 3, "numeric"), "b": (3, 2, "numeric"), "c": (5, 4, "numeric"), "d": (9, 1, "numeric"), } kwargs = { "total_rows": 10, "val_ratio": 0.5, "split_seed": 42, "cache_dir": tmp_path, "source_signature": "synthetic", "prefixes": prefixes, "split_source_namespace": "video", } train = np.load(ensure_video_split_inverse_index(split="train", **kwargs)) val = np.load(ensure_video_split_inverse_index(split="val", **kwargs)) train_mask = train >= 0 val_mask = val >= 0 assert np.all(np.logical_xor(train_mask, val_mask)) assert np.array_equal(train[train_mask], np.arange(train_mask.sum(), dtype=np.int32)) assert np.array_equal(val[val_mask], np.arange(val_mask.sum(), dtype=np.int32)) for start, count, _ in prefixes.values(): assert train_mask[start : start + count].all() or val_mask[start : start + count].all()