"""Data invariants: the renderer must match the spec, and batching must be pure. These are the properties the whole training pipeline rests on - if the renderer drifts from the scene spec, every grounded answer becomes a lie. """ import numpy as np import pytest import torch import omni.render as render import omni.scenes as scenes def test_scene_spec_is_self_consistent(): rng = np.random.default_rng(0) for _ in range(30): sc = scenes.make_scene(rng) assert 1 <= len(sc["objects"]) <= 4 cells = [o["cell"] for o in sc["objects"]] assert len(set(cells)) == len(cells), "one object per grid cell" for o in sc["objects"]: assert o["shape"] in scenes.SHAPES assert o["color"] in scenes.COLORS assert 0.15 <= o["cx"] <= 0.85 and 0.15 <= o["cy"] <= 0.85 def test_numpy_renderer_draws_the_spec(): """The pixel at each object centre must carry that object's colour.""" for seed in range(12): img, _, _, sc = scenes.render_and_ask(seed, seed) for o in sc["objects"]: x, y = int(o["cx"] * 128), int(o["cy"] * 128) got = img[y, x].astype(int) want = np.array(scenes.COLORS[o["color"]]) assert np.abs(got - want).max() <= 60, (o["color"], got, want) @pytest.mark.skipif(not torch.cuda.is_available(), reason="GPU rasteriser") def test_gpu_rasteriser_matches_the_reference(): spec = scenes.spec_batch(list(range(8)), seed=1, px=128) gpu = render.render_specs(spec, px=128, device="cuda", ss=2) assert gpu.shape == (8, 3, 128, 128) and float(gpu.min()) >= 0 and float(gpu.max()) <= 1 for i in range(4): ref = scenes.render_scene(spec["scenes"][i], px=128).astype(np.float32) / 255 mae = float(np.abs(ref - gpu[i].permute(1, 2, 0).cpu().numpy()).mean()) assert mae < 0.02, f"GPU renderer drifted from the reference (MAE {mae:.3f})" def test_answers_are_grounded(): rng = np.random.default_rng(3) seen = 0 for _ in range(80): sc = scenes.make_scene(rng) q, a = scenes.question_answer(sc, rng) assert isinstance(q, str) and isinstance(a, str) and a if q.startswith("how many"): assert a in scenes.NUM_WORD, a seen += 1 assert seen > 5, "the generator should produce counting questions" def test_vocabulary_is_deterministic(): from omni.data import build_vocab a = build_vocab(seed=5, n=200, cache="/tmp/v1.json") b = build_vocab(seed=5, n=200, cache="/tmp/v2.json") assert a == b and len(a) == len(set(a)) from omni.vlm import CombinedVocab import omni.text as t v = CombinedVocab(words=a, phones=t.PHONES) ids = v.words_to_ids(a[:5]) assert v.ids_to_words(ids) == a[:5] assert len(v) == len(v.words) + len(v.phones) def test_batch_indices_are_a_pure_function_of_the_step(): """The resume contract: step N+1 after a restart is the same batch.""" import omni.data as od first = od.batch_indices_for_step(7, 32, seed=0) assert first == od.batch_indices_for_step(7, 32, seed=0) assert first != od.batch_indices_for_step(8, 32, seed=0) assert sorted(first) == sorted(set(first)) straight = [i for s in range(5) for i in od.batch_indices_for_step(s, 8, 0)] resumed = [i for s in range(5) for i in od.batch_indices_for_step(s, 8, 0)] # after a "restart" assert straight == resumed def test_collate_pads_and_masks_the_answer_region(vocab): import omni.data as od spec = scenes.spec_batch(list(range(4)), seed=2, px=64) batch = od.collate_specs(vocab, spec, device="cpu", px=64) ids, targets, mask = batch["ids"], batch["targets"], batch["mask"] assert ids.shape == targets.shape and mask.shape == targets.shape assert bool(mask.any()) and not bool(mask.all()), "loss covers the answer, not the question" assert ids.dtype == torch.long and batch["images"].shape[1:] == (3, 64, 64) for i, start in enumerate(batch["starts"]): assert not bool(mask[i, :max(0, start - 1)].any())