auto-sync 2026-07-02T20:07:13Z workspace (part 9)

workspace/tests/test_tangent_cvae.py

@@ -0,0 +1,59 @@
+from __future__ import annotations
+
+from dovla_cil.generation.tangent_cvae import (
+    build_tangent_cvae_rows,
+    condition_vector,
+    hash_bow,
+)
+
+
+def _target(group_id: str, label: str, value: float) -> dict:
+    return {
+        "group_id": group_id,
+        "task_id": "PickCube-v1",
+        "instruction": "Pick up the cube.",
+        "observation_inline": {"features": [0.1, 0.2, 0.3]},
+        "candidate_type": "near_miss",
+        "tangent_label": label,
+        "delta_utility": 1.0 if label == "positive" else -1.0,
+        "delta_action": [[value, 0.0], [value, 0.0]],
+    }
+
+
+def test_hash_bow_is_unit_scaled_and_deterministic() -> None:
+    left = hash_bow("Pick up the cube", 16)
+    right = hash_bow("Pick up the cube", 16)
+
+    assert left == right
+    assert len(left) == 16
+    assert 0.99 <= sum(value * value for value in left) <= 1.01
+
+
+def test_condition_vector_includes_observation_text_and_task() -> None:
+    task_to_index = {"PickCube-v1": 0, "StackCube-v1": 1}
+    vector = condition_vector(
+        _target("g0", "positive", 0.1),
+        task_to_index=task_to_index,
+        obs_dim=4,
+        text_dim=8,
+    )
+
+    assert len(vector) == 14
+    assert vector[:4] == [0.1, 0.2, 0.3, 0.0]
+    assert vector[-2:] == [1.0, 0.0]
+
+
+def test_build_tangent_cvae_rows_preserves_labels_and_action_dim() -> None:
+    rows, tasks, action_dim = build_tangent_cvae_rows(
+        [
+            _target("g0", "positive", 0.1),
+            _target("g1", "negative", -0.2),
+        ],
+        obs_dim=4,
+        text_dim=8,
+    )
+
+    assert tasks == ["PickCube-v1"]
+    assert action_dim == 4
+    assert [row.example.tangent_label for row in rows] == ["positive", "negative"]
+    assert rows[0].delta == [0.1, 0.0, 0.1, 0.0]

workspace/tests/test_tangent_memory.py

@@ -0,0 +1,69 @@
+from __future__ import annotations
+
+from dovla_cil.generation.tangent_memory import (
+    TangentExample,
+    evaluate_tangent_memory_generator,
+)
+
+
+def _example(
+    group_id: str,
+    label: str,
+    value: float,
+    *,
+    task_id: str = "TaskA",
+    utility: float = 1.0,
+) -> TangentExample:
+    return TangentExample(
+        group_id=group_id,
+        task_id=task_id,
+        candidate_type="near_miss" if label == "positive" else "random_negative",
+        tangent_label=label,
+        delta_utility=utility if label == "positive" else -utility,
+        delta_action=[[value, 0.0], [value, 0.0]],
+    )
+
+
+def test_positive_tangent_memory_reports_support_recall_proxy() -> None:
+    examples = [
+        _example("train-a", "positive", 0.1),
+        _example("train-b", "positive", 0.4),
+        _example("val-a", "positive", 0.11),
+        _example("val-a", "negative", -0.6),
+        _example("val-b", "positive", 0.39),
+    ]
+
+    report = evaluate_tangent_memory_generator(
+        examples,
+        k_values=(1, 2),
+        thresholds=(0.05,),
+        val_fraction=0.5,
+        seed=4,
+        diversity_weight=0.0,
+    )
+
+    assert report["overall"]["num_groups"] >= 1
+    assert 0.0 <= report["overall"]["ptr_proxy_at_2_thr_0p05"] <= 1.0
+    assert report["proposal_count_by_task"]["TaskA"] == 2
+
+
+def test_positive_tangent_memory_uses_train_positive_prototypes_only() -> None:
+    examples = [
+        _example("train-b", "positive", 10.0, utility=1.0),
+        _example("val-a", "positive", 0.0, utility=100.0),
+        _example("val-a", "negative", 10.0, utility=1.0),
+    ]
+
+    report = evaluate_tangent_memory_generator(
+        examples,
+        k_values=(1,),
+        thresholds=(0.1,),
+        val_fraction=0.5,
+        seed=0,
+        diversity_weight=0.0,
+    )
+
+    assert report["train_positive_by_task"]["TaskA"] == 1
+    assert report["proposal_count_by_task"]["TaskA"] == 1
+    assert report["overall"]["ptr_proxy_at_1_thr_0p1"] == 0.0
+    assert report["groups"][0]["positive_min_rms_l2_at_1"] > 1.0