| """Select the integration checkpoint by the broader V2 validation split only.""" |
| import json |
| from pathlib import Path |
|
|
| import torch |
|
|
|
|
| def main(): |
| root = Path(__file__).resolve().parent |
| candidates = [] |
| for report,artifact,name in ( |
| ("ood_v1_v2.json","results/best.pt","Frozen V1 raw MLP"), |
| ("ood_centered_v2.json","artifacts_centered/best.pt","Frozen V1 centered MLP"), |
| ("results_v2_raw/results.json","artifacts_v2_raw/best.pt","V2 raw balanced MLP"), |
| ("results_v2_centered/results.json","artifacts_v2_centered/best.pt","V2 centered balanced MLP"), |
| ("results_v2_gru/results.json","artifacts_v2_gru/best.pt","V2 centered balanced GRU")): |
| if not (root/report).exists(): continue |
| payload = json.loads((root/report).read_text()) |
| validation = payload.get("validation",payload.get("selected",{}).get("validation")) |
| candidates.append({"name":name,"report":report,"artifact":artifact,"validation":validation,"test":payload["test"]}) |
| if len(candidates)!=5: raise ValueError("All five predefined candidate comparisons must finish before selecting") |
| chosen = min(candidates,key=lambda c:c["validation"]["eer"]) |
| state = torch.load(root/chosen["artifact"],map_location="cpu",weights_only=False) |
| state["selection_validation"] = chosen["validation"] |
| state["test"] = chosen["test"] |
| state["selection_protocol"] = "Minimum V2 validation EER among five predefined candidates; test does not choose weights or thresholds" |
| destination = root/"artifacts" |
| destination.mkdir(exist_ok=True) |
| torch.save(state,destination/"best.pt") |
| params = sum(v.numel() for k,v in state["model"].items() if k not in ("teacher_mean","teacher_radius")) |
| selection = {"checkpoint":"best.pt","name":chosen["name"],"config":state["config"], |
| "parameters":params,"verification_threshold":state["verification_threshold"], |
| "activity_threshold":state["activity_threshold"],"recommended_enabled":False,"pilot_pass":False, |
| "validation":chosen["validation"],"test":chosen["test"], |
| "selection_protocol":state["selection_protocol"], |
| "threshold_source":"Frozen selected checkpoint's original validation-only calibration; no test threshold adjustment", |
| "activity_calibration_note":"Speech activity probability/context coverage is not speaker identity probability. Original validation nonspeech scores calibrate abstention, microphone RMS adds a separate silence gate.", |
| "note":"Experimental synthetic-voice pilot: insufficient held-out verification and overlap robustness. Keep default identity gate OFF; do not use for authentication. Full-scale training was not started because the pilot failed reliability checks.", |
| "candidates":[{"name":c["name"],"validation_eer":c["validation"]["eer"],"report":c["report"]} for c in candidates], |
| "feature_source":state["feature_source"],"teacher":state["teacher"]} |
| folder = {"Frozen V1 raw MLP":"results","Frozen V1 centered MLP":"results_centered", |
| "V2 raw balanced MLP":"results_v2_raw","V2 centered balanced MLP":"results_v2_centered", |
| "V2 centered balanced GRU":"results_v2_gru"}[chosen["name"]] |
| original = json.loads((root/folder/"results.json").read_text()) |
| selection["cpu_batch1_ms"] = original["selected"]["cpu_batch1_ms"] |
| (destination/"selection.json").write_text(json.dumps(selection,indent=2)) |
| print(json.dumps({"selected":chosen["name"],"validation_eer":chosen["validation"]["eer"],"test_eer":chosen["test"]["eer"],"recommended_enabled":False})) |
|
|
|
|
| if __name__ == "__main__": main() |
|
|
