import importlib.util from pathlib import Path import sys import pytest import torch ROOT = Path(__file__).resolve().parents[1] def load_script(name): sys.path.insert(0, str(ROOT / "scripts/data")) spec = importlib.util.spec_from_file_location(name, ROOT / "scripts/data" / f"{name}.py") module = importlib.util.module_from_spec(spec);spec.loader.exec_module(module) return module def test_vqa_candidate_gate_requires_nine_human_votes_and_visual_question(): m = load_script("prepare_swa_vqav2") questions = [{"question_id": i, "image_id": i, "question": q} for i, q in enumerate( ["What color is this shirt?", "What color is this shirt?", "Why is she wearing this shirt?"])] annotations = [{"question_id": i, "image_id": i, "multiple_choice_answer": "red", "answer_type": "other", "answers": [{"answer": "red"}] * n + [{"answer": "blue"}] * (10 - n)} for i, n in enumerate([9, 8, 10])] rows, rejected = m.select(questions, annotations, split="train", limit=10) assert [r["question_id"] for r in rows] == [0] assert rejected["human_agreement_below_threshold"] == 1 assert rejected["question_not_conservative_visible_state"] == 1 assert not m.conservative_question("Can this boy get injured easily?") assert not m.conservative_question("Was this taken in the United States?") assert not m.conservative_question("Is this scene in Paris?") assert m.conservative_question("What is the horse doing?") def test_speech_gate_does_not_hide_omitted_words(): m = load_script("encode_swa_vqav2") assert m.canonical_speech("Are there two dogs?") == m.canonical_speech("Are there 2 dogs.") assert m.canonical_speech("Is the dog red?") != m.canonical_speech("Is dog red?") def test_camera_packet_grouping_causal_and_same_cadence(): m = load_script("encode_swa_vqav2") available, groups = m.camera_groups(45) assert groups[0] == [0, 1, 2, 3, 4] assert set(map(len, groups)) == {3, 4, 5} flat = [i for group in groups for i in group] assert flat == list(range(len(flat))) assert all(max(group) / 8 <= deadline + 1e-6 for group, deadline in zip(groups, available)) def test_factorized_packets_reconstruct_without_mutating_shared_bases(tmp_path): from psi.data.seed.swa_frontend import load_swa_frontend packet = {"embeds": torch.tensor([[0., 0.], [1., 1.]]), "visual_mask": torch.tensor([True, False]), "audio_mask": torch.tensor([False, True]), "camera_indices": [0, 1, 2], "answer_target": ""} base = tmp_path / "base.pt";visual = tmp_path / "visual.pt";front = tmp_path / "front.pt" torch.save({"packets": [packet, packet], "system": "test"}, base) torch.save({"3": torch.tensor([[3., 3.]])}, visual) torch.save({"format": "static_factorized_v1", "base_frontend_path": str(base), "visual_features_path": str(visual), "audio_patches": {1: torch.tensor([[4., 4.]])}, "reply_windows": [1], "answer_target": "yes", "timeline": {}}, front) first = load_swa_frontend(front)["packets"] assert first[0]["embeds"].tolist() == [[3., 3.], [1., 1.]] assert first[1]["embeds"].tolist() == [[3., 3.], [4., 4.]] assert first[0]["answer_target"] == "" and first[1]["answer_target"] == "yes" first[0]["embeds"].zero_() assert load_swa_frontend(front)["packets"][0]["embeds"].tolist() == [[3., 3.], [1., 1.]] def test_swa_eval_indices_cover_entire_manifest_without_repeats(): from psi.trainers.swa_validation import validation_indices assert validation_indices(0) == [] assert validation_indices(3) == [0, 1, 2] indices = validation_indices(1000) assert len(indices) == len(set(indices)) == 16 assert indices[-1] > 900