| import json |
| import subprocess |
| from pathlib import Path |
|
|
| OUTPUT = Path("/app/output.json") |
| ROOT = Path("/app") |
|
|
|
|
| def load_output(): |
| assert OUTPUT.is_file() |
| return json.loads(OUTPUT.read_text()) |
|
|
|
|
| def run_reference(): |
| result = subprocess.run( |
| ["python3", "/tests/fixtures/regression_probe.py", "/app", "/tmp/recomputed.json"], |
| cwd="/tests", capture_output=True, text=True, timeout=30, |
| ) |
| assert result.returncode == 0, result.stderr |
| return json.loads(Path("/tmp/recomputed.json").read_text()) |
|
|
|
|
| def test_gate_schema_and_files(): |
| payload = load_output() |
| assert payload["output_schema_version"] == "tbench.early_stopping.repair.v1" |
| assert set(payload["cases"]) == {"deferred_patience", "post_min_epochs", "min_steps_only", "recovery_after_defer"} |
| assert set(payload["fit_loop_gates"]) == {"epoch_pending", "step_pending", "all_met", "no_minimums"} |
| assert (ROOT / "vendor/pytorch-lightning/src/lightning/pytorch/callbacks/early_stopping.py").is_file() |
| assert (ROOT / "vendor/pytorch-lightning/src/lightning/pytorch/loops/fit_loop.py").is_file() |
| assert (ROOT / "run_regression.py").is_file() |
|
|
|
|
| def test_core_independent_recompute(): |
| expected = json.loads((Path("/tests/fixtures/expected.json")).read_text()) |
| assert load_output() == expected |
| assert run_reference() == expected |
|
|
|
|
| def test_core_deferred_and_ordinary_stop(): |
| payload = load_output()["summary"] |
| assert payload["deferred_stop"] is False |
| assert payload["ordinary_stop"] is True |
| assert payload["recovered"] is True |
|
|
|
|
| def test_core_min_steps_boundary(): |
| case = load_output()["cases"]["min_steps_only"] |
| assert case[-1]["should_stop"] is True |
| assert case[-1]["stopping_reason"] == "PATIENCE_EXHAUSTED" |
| assert case[-1]["stopped_epoch"] == 1 |
|
|
|
|
| def test_core_fit_loop_gate_matrix(): |
| gates = load_output()["fit_loop_gates"] |
| assert gates["epoch_pending"] == {"met_min_epochs": False, "met_min_steps": True, "can_stop_early": False} |
| assert gates["step_pending"] == {"met_min_epochs": True, "met_min_steps": False, "can_stop_early": False} |
| assert gates["all_met"] == {"met_min_epochs": True, "met_min_steps": True, "can_stop_early": True} |
| assert gates["no_minimums"] == {"met_min_epochs": True, "met_min_steps": True, "can_stop_early": True} |
|
|
|
|
| def test_edge_source_contract(): |
| callback = (ROOT / "vendor/pytorch-lightning/src/lightning/pytorch/callbacks/early_stopping.py").read_text() |
| fit_loop = (ROOT / "vendor/pytorch-lightning/src/lightning/pytorch/loops/fit_loop.py").read_text() |
| assert "trainer.strategy.reduce_boolean_decision" in callback |
| assert "trainer.fit_loop._met_min_epochs" in callback |
| assert "self.stopping_reason = EarlyStoppingReason.NOT_STOPPED" in callback |
| assert "def _met_min_epochs" in fit_loop |
| assert "def _met_min_steps" in fit_loop |
| assert "return self._met_min_epochs and self._met_min_steps" in fit_loop |
|
|
