| """CPU tests of the scene-specification, render and build helpers (no models). |
| """ |
| import csv, json, pathlib, subprocess, sys |
|
|
| import numpy as np |
| import pytest |
|
|
| from strangebench.generation import coco_class_prompts as ccp |
| from strangebench.scene_specs import layout_solver as ls |
| from strangebench.audit.audit_diffusion_scenes import dedup_cross_class |
| from strangebench.generation import gen_sd35l_e2e as gen |
|
|
| ROOT = pathlib.Path(__file__).resolve().parents[1] |
|
|
|
|
| def test_choose_attempt_first_complete_wins(): |
| assert gen.choose_attempt([2, 3, 3], 3) == 1 |
| assert gen.choose_attempt([3], 3) == 0 |
| assert gen.choose_attempt([1, 4, 3], 3) == 1 |
|
|
|
|
| def test_choose_attempt_most_found_first_on_ties(): |
| assert gen.choose_attempt([1, 2, 2, 1], 3) == 1 |
| assert gen.choose_attempt([0, 0], 2) == 0 |
| with pytest.raises(ValueError): |
| gen.choose_attempt([], 2) |
|
|
|
|
| CATS = {1: "a", 2: "b"} |
|
|
|
|
| def _ann(aid, cid, bbox, score): |
| return dict(id=aid, image_id=1, category_id=cid, bbox=bbox, score=score) |
|
|
|
|
| def test_dedup_keeps_intended_class_drops_other(): |
| anns = [_ann(1, 1, [0, 0, 100, 100], 0.4), _ann(2, 2, [0, 0, 100, 90], 0.95)] |
| dropped, flag = dedup_cross_class(anns, {"a"}, CATS, iou_thr=0.5) |
| assert [d["ann_id"] for d in dropped] == [2] |
| assert dropped[0]["kept_ann_id"] == 1 and dropped[0]["kept_class"] == "a" |
| assert dropped[0]["iou"] == pytest.approx(0.9, abs=1e-3) |
| assert dropped[0]["dropped_intended"] is False and flag is False |
|
|
|
|
| def test_dedup_both_intended_keeps_higher_score_and_flags_scene(): |
| anns = [_ann(1, 1, [0, 0, 100, 100], 0.6), _ann(2, 2, [0, 0, 100, 90], 0.8)] |
| dropped, flag = dedup_cross_class(anns, {"a", "b"}, CATS, iou_thr=0.5) |
| assert [d["ann_id"] for d in dropped] == [1] and flag is True |
|
|
|
|
| def test_dedup_ignores_low_iou_same_class_and_off(): |
| far = [_ann(1, 1, [0, 0, 100, 100], 0.6), _ann(2, 2, [500, 500, 100, 100], 0.8)] |
| assert dedup_cross_class(far, {"a", "b"}, CATS, 0.5) == ([], False) |
| same = [_ann(1, 1, [0, 0, 100, 100], 0.6), _ann(2, 1, [0, 0, 100, 90], 0.8)] |
| assert dedup_cross_class(same, {"a"}, CATS, 0.5) == ([], False) |
| anns = [_ann(1, 1, [0, 0, 100, 100], 0.6), _ann(2, 2, [0, 0, 100, 90], 0.8)] |
| assert dedup_cross_class(anns, {"a"}, CATS, 0.0) == ([], False) |
|
|
|
|
| def test_size_word_thresholds(): |
| lo, hi = ccp.SIZE_WORD_THRESHOLDS |
| assert ccp.size_word(lo / 2) == "small" |
| assert ccp.size_word(lo) == "medium" and ccp.size_word((lo + hi) / 2) == "medium" |
| assert ccp.size_word(hi) == "large" and ccp.size_word(0.9) == "large" |
| assert ccp.rel_box_area([0, 0, 512, 512], 1024) == pytest.approx(0.25) |
| assert ccp.size_word_for_box([24, 24, 1000, 1000]) == "large" |
| assert ccp.size_word_for_box([0, 0, 100, 100]) == "small" |
|
|
|
|
| SCENE = dict(scene_id="tier2_1atypical__a064_s0000", sub_tier="tier2_1atypical", n_partners=1, |
| labels=["mouse", "cow"], layout_boxes=[[100, 100, 300, 300], [700, 700, 1000, 1000]], |
| img_size=1024, prompt="A photorealistic photo of a mouse and a cow, in an unusual setting") |
|
|
|
|
| def _v1_objects_block(scene): |
| return "\n".join(f" - {l} → {gen.bbox_to_position(bb)} of the image → {ccp.describe(l)}" |
| for l, bb in zip(scene["labels"], scene["layout_boxes"])) |
|
|
|
|
| def test_display_name_override(): |
| assert ccp.display_name("mouse") == "computer mouse" |
| assert ccp.display_name("keyboard") == "computer keyboard" |
| assert ccp.display_name("mouse", overrides=False) == "mouse" |
| assert ccp.display_name("cow") == "cow" |
| block = gen.make_objects_block(SCENE) |
| assert block.startswith(" - computer mouse → ") and " - cow → " in block |
| assert gen.caption_names(["mouse", "cow"]) == ["computer mouse", "cow"] |
| assert gen.caption_names(["mouse", "cow"], display_names=False) == ["mouse", "cow"] |
| fixed = gen.apply_display_names(SCENE["prompt"], SCENE["labels"]) |
| assert fixed == "A photorealistic photo of a computer mouse and a cow, in an unusual setting" |
| assert gen.apply_display_names(fixed, SCENE["labels"]) == fixed |
|
|
|
|
| def test_v1_wording_byte_identical_when_overrides_and_visibility_off(): |
| ctx, phrase = gen.tier_descriptor(SCENE) |
| expected = gen.SPATIAL_PROMPT_TEMPLATE.format(context_label=ctx, tier_phrase=phrase, |
| objects_block=_v1_objects_block(SCENE)) |
| assert gen.build_user_prompt(SCENE, visibility=False, display_names=False) == expected |
| assert gen.make_objects_block(SCENE, display_names=False) == _v1_objects_block(SCENE) |
| assert " - mouse → " in expected and " - computer mouse → " not in expected |
|
|
|
|
| def test_visibility_prompt_adds_bullet_and_size_words(): |
| p = gen.build_user_prompt(SCENE, visibility=True) |
| assert "FULLY VISIBLE and unoccluded" in p and "size in the frame" in p |
| assert " - computer mouse → top left of the image → small in the frame → " in p |
| assert " - cow → bottom right of the image → medium in the frame → " in p |
| assert gen.SPATIAL_PROMPT_TEMPLATE_VISIBILITY.endswith(gen.VISIBILITY_BULLET) |
| assert "\\" not in p and p.count("\n- ") == 7 |
| assert gen.SPATIAL_PROMPT_TEMPLATE not in (gen.SPATIAL_PROMPT_TEMPLATE_VISIBILITY,) |
| v1 = gen.build_user_prompt(SCENE, visibility=False) |
| assert "in the frame" not in v1 |
|
|
|
|
| def test_sam3_prompts_for(): |
| prim = {"keyboard": "a computer keyboard", "bed": "a picture of a {category}"} |
| sec = {"keyboard": "a keyboard"} |
| assert gen.sam3_prompts_for("cow", prim, sec) == ["a cow"] |
| assert gen.sam3_prompts_for("cow", prim, sec, union=True) == ["a cow", "cow"] |
| assert gen.sam3_prompts_for("keyboard", prim, sec, union=True) == ["a computer keyboard", "a keyboard"] |
| assert gen.sam3_prompts_for("bed", prim, sec, union=True) == ["a picture of a bed", "bed"] |
|
|
|
|
| def _mask(x0, x1): |
| m = np.zeros((64, 64), dtype=np.uint8) |
| m[16:48, x0:x1] = 255 |
| return m |
|
|
|
|
| def test_merge_prompt_results(): |
| assert gen.merge_prompt_results([]) is None |
| same, same2, other = _mask(10, 40), _mask(12, 42), _mask(50, 60) |
| m, s, hits = gen.merge_prompt_results([(same, 0.7, "a cow"), (same2, 0.9, "cow")], iou_thr=0.5) |
| assert s == 0.9 and hits == ["cow", "a cow"] |
| assert m[20, 11] == 255 and m[20, 41] == 255 |
| m, s, hits = gen.merge_prompt_results([(same, 0.7, "a cow"), (other, 0.6, "cow")], iou_thr=0.5) |
| assert s == 0.7 and hits == ["a cow"] and m[20, 55] == 0 |
| m, s, hits = gen.merge_prompt_results([(same, 0.7, "a cow"), (same2, 0.7, "cow")]) |
| assert hits[0] == "a cow" |
|
|
|
|
| K6 = ["person", "dog", "bench", "kite", "bottle", "airplane"] |
|
|
|
|
| @pytest.fixture(scope="module") |
| def priors(): |
| assert ls.DEFAULT_PRIORS_PATH.exists(), ls.DEFAULT_PRIORS_PATH |
| p = ls.load_size_priors() |
| assert len(p) == 80 |
| return p |
|
|
|
|
| @pytest.mark.parametrize("area_scale", [1.0, 3.0]) |
| def test_solver_k6_non_overlap_and_margins(priors, area_scale): |
| cfg = ls.SolverConfig() |
| for seed in range(10): |
| info = ls.solve_layout(K6, priors, 1024, seed=seed, area_scale=area_scale) |
| assert info["layout_source"] == "solver", (seed, info) |
| boxes = info["layout_boxes"] |
| assert len(boxes) == 6 and ls.check_layout(boxes, K6) == [] |
| for x1, y1, x2, y2 in boxes: |
| assert cfg.margin <= x1 < x2 <= 1024 - cfg.margin and cfg.margin <= y1 < y2 <= 1024 - cfg.margin |
| assert min(x2 - x1, y2 - y1) >= cfg.min_short_side |
| assert (x2 - x1) * (y2 - y1) <= cfg.max_rel_area * 1024 * 1024 + 1 |
| for i in range(6): |
| for j in range(i + 1, 6): |
| assert ls.box_iou(boxes[i], boxes[j]) <= cfg.max_iou |
| for lbl, b in zip(K6, boxes): |
| if lbl in ls.GROUND_CONTACT_CLASSES: |
| assert b[3] >= 0.4 * 1024 |
| a = ls.solve_layout(K6, priors, 1024, seed=7, area_scale=area_scale) |
| assert a == ls.solve_layout(K6, priors, 1024, seed=7, area_scale=area_scale) |
| assert a["layout_boxes"] != ls.solve_layout(K6, priors, 1024, seed=8, area_scale=area_scale)["layout_boxes"] |
|
|
|
|
| def test_stage1_solver_seed_is_process_independent(): |
| from strangebench.scene_specs.select_scenes_and_layouts import solver_seed, solver_layout |
| assert solver_seed(42, "tier5_4typical__a001_s0000") == solver_seed(42, "tier5_4typical__a001_s0000") |
| assert solver_seed(42, "x") != solver_seed(43, "x") |
| info = solver_layout(K6, 1024, {}, seed=1, area_scale=3.0) |
| assert info["layout_source"] == "solver" and info["layout_attempts"] == 0 |
| assert info["solver_attempts"] > 0 and info["solver_area_scale"] == 3.0 |
| fb = solver_layout(K6, 1024, {}, seed=1, area_scale=3.0, source="solver_fallback", llm_attempts=3) |
| assert fb["layout_source"] == "solver_fallback" and fb["layout_attempts"] == 3 |
|
|
|
|
| def _run(args, cwd): |
| r = subprocess.run([sys.executable] + args, capture_output=True, text=True, cwd=str(cwd)) |
| assert r.returncode == 0, r.stdout[-2000:] + r.stderr[-2000:] |
| return r |
|
|
|
|
| def test_fragment_to_manifest_lever_columns(tmp_path): |
| from PIL import Image |
| raw = tmp_path / "raw" |
| tier = raw / "tier2_1atypical" |
| (tier / "raw").mkdir(parents=True) |
| Image.new("RGB", (8, 8)).save(tier / "raw" / "v2_s1.png") |
| Image.new("RGB", (8, 8)).save(tier / "raw" / "v2_s2.png") |
| base = dict(sub_tier="tier2_1atypical", labels=["mouse", "cow"]) |
| (tier / "v2_s1_anns.json").write_text(json.dumps(dict(base, scene_id="v2_s1", |
| annotations=[dict(category_name="mouse", bbox=[1, 1, 3, 3], score=0.9), |
| dict(category_name="cow", bbox=[4, 4, 3, 3], score=0.8)], |
| render_attempts=2, attempt_seeds=[10, 11], chosen_attempt=1, complete_at_render=True, |
| n_found_per_attempt=[1, 2], visibility_prompt=True, sam3_prompt_union=False))) |
| (tier / "v2_s2_anns.json").write_text(json.dumps(dict(base, scene_id="v2_s2", |
| annotations=[dict(category_name="cow", bbox=[4, 4, 3, 3], score=0.8)]))) |
| cats = tmp_path / "categories.json" |
| cats.write_text(json.dumps(dict(idx_to_name={"0": "mouse", "1": "cow"}, id_to_idx={"74": 0, "21": 1}))) |
| _run(["strangebench/audit/merge_per_scene_anns.py", "--out-root", str(raw), "--categories-json", str(cats)], ROOT) |
| coco = json.load(open(tier / "annotations.json")) |
| by_sid = {im["scene_id"]: im for im in coco["images"]} |
| assert by_sid["v2_s1"]["render_attempts"] == 2 and by_sid["v2_s1"]["visibility_prompt"] is True |
| assert by_sid["v2_s1"]["complete_at_render"] is True and by_sid["v2_s1"]["attempt_seeds"] == [10, 11] |
| assert "render_attempts" not in by_sid["v2_s2"] |
|
|
| clean = tmp_path / "clean" / "tier2_1atypical" |
| (clean / "images").mkdir(parents=True) |
| for sid in ("v2_s1", "v2_s2"): |
| (clean / "images" / f"{sid}.png").write_bytes((tier / "raw" / f"{sid}.png").read_bytes()) |
| json.dump(coco, open(clean / "annotations.json", "w")) |
| specs = tmp_path / "scene_specs.json" |
| json.dump([dict(scene_id="v2_s1", anchor={"idx": 64, "name": "mouse"}, partners=[{"idx": 19, "name": "cow"}], |
| sub_tier="tier2_1atypical", layout_source="solver", variant="k56"), |
| dict(scene_id="v2_s2", anchor={"idx": 64, "name": "mouse"}, partners=[{"idx": 19, "name": "cow"}], |
| sub_tier="tier2_1atypical", layout_source="qwen")], open(specs, "w")) |
| dst = tmp_path / "v2" |
| _run(["strangebench/strict/build_dataset.py", "--source", "v2", str(tmp_path / "clean"), str(specs), |
| "--dst", str(dst)], ROOT) |
| rows = {r["scene_id"]: r for r in csv.DictReader(open(dst / "tier_manifest.csv"))} |
| assert set(rows) == {"v2_s1", "v2_s2"} |
| assert rows["v2_s1"]["variant"] == "k56" and rows["v2_s1"]["layout_source"] == "solver" |
| assert rows["v2_s1"]["render_attempts"] == "2" and rows["v2_s1"]["visibility_prompt"] == "1" |
| assert rows["v2_s2"]["layout_source"] == "qwen" and rows["v2_s2"]["variant"] == "" |
| assert rows["v2_s2"]["render_attempts"] == "" and rows["v2_s2"]["visibility_prompt"] == "" |
| assert rows["v2_s1"]["complete"] == "1" and rows["v2_s2"]["complete"] == "0" |
|
|
