GotthardGandClaude Fable 5.1
62f568b0f6
The cluster's first run was gated against a baseline measured on the Spark. The same weights score lower there (F1 0.840 vs 0.861, mask score 0.706 vs 0.757 on the 203 gold frames), so the gate failed a model that beats the active one on F1 in 98.7 % of bootstrap draws when both are measured on one machine. Gold reports now record the environment (ultralytics, torch, host). When the active model's report comes from other versions, train_run.py measures its weights again in the job and gates against that; the predictions feed the paired bootstrap. Without the weights there is no verdict and no promotion. The SLURM scripts pull the active model's weights. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
111 lines
6.1 KiB
Python
111 lines
6.1 KiB
Python
"""tools/train_run.py gate: recall-aware mask score and the pass rule."""
|
|
import json
|
|
|
|
from tools import train_run
|
|
|
|
|
|
def _report(f1, iou, tp, fn, classes=None):
|
|
return {"overall": {"f1": f1, "mean_mask_iou": iou, "tp": tp, "fn": fn, "fp": 0, "precision": 1.0, "recall": tp / (tp + fn)},
|
|
"per_class": classes or {}}
|
|
|
|
|
|
def test_mask_score_weights_by_recall():
|
|
assert abs(train_run.mask_score({"mean_mask_iou": 0.9, "tp": 60, "fn": 40}) - 0.54) < 1e-9
|
|
assert train_run.mask_score({"mean_mask_iou": 0.9, "tp": 0, "fn": 0}) == 0.0
|
|
|
|
|
|
def test_gate_prefers_finding_more_with_slightly_looser_masks(tmp_path):
|
|
base = tmp_path / "base.json"
|
|
base.write_text(json.dumps(_report(0.69, 0.903, 60, 40, {"loop_all": {"f1": 0.71}, "pin": {"f1": 0.84}})))
|
|
new = _report(0.85, 0.883, 85, 15, {"loop_all": {"f1": 0.90}, "pin": {"f1": 0.92}})
|
|
g = train_run.gate(new, base)
|
|
assert g["pass"] is True and g["mask_iou_delta"] > 0 and g["matched_only_iou_delta"] < 0
|
|
|
|
|
|
def test_gate_refuses_a_guarded_class_regression(tmp_path):
|
|
base = tmp_path / "base.json"
|
|
base.write_text(json.dumps(_report(0.70, 0.90, 60, 40, {"loop_all": {"f1": 0.80}, "pin": {"f1": 0.84}})))
|
|
new = _report(0.80, 0.90, 80, 20, {"loop_all": {"f1": 0.75}, "pin": {"f1": 0.90}})
|
|
assert train_run.gate(new, base)["pass"] is False
|
|
|
|
|
|
def test_no_baseline_means_no_gate():
|
|
assert train_run.gate(_report(0.8, 0.9, 1, 0), None) is None
|
|
|
|
|
|
def _frames(per_frame_tp, fp=0, objects=2, iou=0.9):
|
|
"""One class, ``objects`` gold objects per frame, ``tp`` of them found."""
|
|
return {f"f{i}": {0: (float(tp), float(fp), float(objects - tp), iou * tp)} for i, tp in enumerate(per_frame_tp)}
|
|
|
|
|
|
def test_paired_bootstrap_separates_a_real_gain_from_a_tie():
|
|
from tools import gate_stats
|
|
base = _frames([1] * 60)
|
|
better = _frames([2] * 40 + [1] * 20) # finds more on two thirds of the frames
|
|
same = _frames([1] * 60)
|
|
r = gate_stats.paired_bootstrap(better, base, draws=400)
|
|
assert r["frames"] == 60 and r["f1"]["win"] > 0.99 and r["f1"]["lo"] > 0 and r["mask_score"]["win"] > 0.99
|
|
t = gate_stats.paired_bootstrap(same, base, draws=400)
|
|
assert t["f1"]["delta"] == 0 and t["f1"]["win"] == 0.0 # never strictly better
|
|
|
|
|
|
def test_paired_bootstrap_pairs_only_on_common_frames():
|
|
from tools import gate_stats
|
|
cand = _frames([2] * 10); base = _frames([1] * 6)
|
|
assert gate_stats.paired_bootstrap(cand, base, draws=50)["frames"] == 6
|
|
assert gate_stats.paired_bootstrap(cand, {}, draws=50) == {"frames": 0, "draws": 0}
|
|
|
|
|
|
def test_gate_with_bootstrap_needs_a_clear_win(tmp_path):
|
|
base = tmp_path / "base.json"
|
|
base.write_text(json.dumps(_report(0.80, 0.90, 80, 20, {"loop_all": {"f1": 0.90}, "pin": {"f1": 0.90}})))
|
|
new = _report(0.806, 0.90, 81, 19, {"loop_all": {"f1": 0.90}, "pin": {"f1": 0.90}})
|
|
noisy = {"frames": 150, "draws": 2000, "f1": {"delta": 0.006, "lo": -0.02, "hi": 0.03, "win": 0.66},
|
|
"mask_score": {"delta": 0.004, "lo": -0.02, "hi": 0.03, "win": 0.60}}
|
|
g = train_run.gate(new, base, noisy)
|
|
assert g["pass_point_rule"] is True and g["pass"] is False and g["rule"] == "paired bootstrap"
|
|
clear = {"frames": 150, "draws": 2000, "f1": {"delta": 0.05, "lo": 0.02, "hi": 0.08, "win": 1.0},
|
|
"mask_score": {"delta": 0.04, "lo": 0.01, "hi": 0.07, "win": 0.99}}
|
|
assert train_run.gate(new, base, clear)["pass"] is True
|
|
lopsided = {"frames": 150, "draws": 2000, "f1": {"delta": 0.03, "lo": 0.0, "hi": 0.06, "win": 0.97},
|
|
"mask_score": {"delta": -0.04, "lo": -0.07, "hi": -0.01, "win": 0.01}}
|
|
assert train_run.gate(new, base, lopsided)["pass"] is False # wins F1 but clearly loses the masks
|
|
|
|
|
|
def test_gate_without_predictions_keeps_the_point_rule(tmp_path):
|
|
base = tmp_path / "base.json"
|
|
base.write_text(json.dumps(_report(0.69, 0.903, 60, 40)))
|
|
g = train_run.gate(_report(0.85, 0.883, 85, 15), base)
|
|
assert g["rule"] == "point" and g["pass"] is True and g["pass_point_rule"] is True
|
|
|
|
|
|
def test_same_environment_trusts_legacy_reports_only_on_the_spark():
|
|
here = {"host": "gpu102", "ultralytics": "8.4.37", "torch": "2.5.1"}
|
|
assert train_run.same_environment({"ultralytics": "8.4.37", "torch": "2.5.1", "host": "mx-spark-01"}, here) is True
|
|
assert train_run.same_environment({"ultralytics": "8.4.160", "torch": "2.5.1"}, here) is False
|
|
assert train_run.same_environment(None, here) is False # legacy report, read on the cluster
|
|
assert train_run.same_environment(None, {"host": "mx-spark-01", "ultralytics": "x", "torch": "y"}) is True
|
|
|
|
|
|
def test_baseline_from_another_environment_is_measured_again_or_refused(tmp_path, monkeypatch):
|
|
reg = tmp_path / "models" / "registry" / "v-old"; reg.mkdir(parents=True)
|
|
stored = reg / "gold_report.json"
|
|
stored.write_text(json.dumps(dict(_report(0.84, 0.865, 80, 20), environment={"ultralytics": "8.4.160", "torch": "2.14.0"})))
|
|
monkeypatch.setattr(train_run, "ROOT", tmp_path)
|
|
monkeypatch.setattr(train_run, "environment", lambda: {"host": "gpu102", "ultralytics": "8.4.37", "torch": "2.5.1"})
|
|
# no weights here: no baseline to gate against, and a reason
|
|
path, note = train_run.baseline_here(stored, ["loop_all"], 0.25)
|
|
assert path is None and "weights are not here" in note
|
|
# an lfs pointer is not weights either
|
|
(reg / "weights.pt").write_text("version https://git-lfs.github.com/spec/v1")
|
|
assert train_run.baseline_here(stored, ["loop_all"], 0.25)[0] is None
|
|
# real weights: measured again here, the local report is the baseline
|
|
(reg / "weights.pt").write_bytes(b"0" * 1_100_000)
|
|
monkeypatch.setattr(train_run, "measure", lambda w, work, names, conf=0.25: _report(0.86, 0.886, 85, 15))
|
|
path, note = train_run.baseline_here(stored, ["loop_all"], 0.25)
|
|
assert path.name == "gold_report.local.json" and json.loads(path.read_text())["overall"]["f1"] == 0.86
|
|
assert "measured again" in note
|
|
# same environment: the stored report is used as it is
|
|
monkeypatch.setattr(train_run, "environment", lambda: {"host": "x", "ultralytics": "8.4.160", "torch": "2.14.0"})
|
|
assert train_run.baseline_here(stored, ["loop_all"], 0.25) == (stored, None)
|