| |
| |
| """Write assets/nn_reference.npz: the self-test's outputs in the reference environment, for the near-identical check. |
| |
| source env/canonical.sh && python -s tools/make_nn_reference.py --device cuda:0 |
| |
| Runs the start-up self-test's cases (nn/torch_backend.py, golden_cases) in the exact lane and keeps their outputs only |
| if every one's digest equals the release's (assets/nn_golden.json), so the arrays are the release's own outputs. |
| Outside the reference environment, a Diarizer that is not strict compares its self-test outputs with these |
| (torch_backend.NEAR). Stores segment_b8 (the smaller segmentation batches are its first rows) and the voiceprint cases. |
| """ |
| import argparse |
| import hashlib |
| import io |
| import json |
| import sys |
| from pathlib import Path |
|
|
| import numpy as np |
|
|
| PROJ = Path(__file__).resolve().parents[1] |
|
|
|
|
| def main(): |
| ap = argparse.ArgumentParser() |
| ap.add_argument("--device", default="cuda:0") |
| ap.add_argument("--out", type=Path, default=PROJ / "assets/nn_reference.npz") |
| a = ap.parse_args() |
| sys.path.insert(0, str(PROJ / "python")) |
| from anuvani_diar.nn.torch_backend import TorchNN, _digest, golden_cases |
| golden = PROJ / "assets/nn_golden.json" |
| nn = TorchNN(PROJ / "artifacts/bundle", devices=[a.device], golden=golden, strict=True) |
| want = json.loads(golden.read_text())["digests"] |
| nn._force_exact = True |
| out = {name: fn() for name, fn in golden_cases(nn, 0)} |
| bad = [k for k in want if _digest(out[k]) != want[k]] |
| if bad: |
| raise SystemExit(f"not the release's outputs: {bad}") |
| keep = {k: np.ascontiguousarray(v, dtype=np.float32) for k, v in out.items() if not k.startswith("segment") or k == "segment_b8"} |
| buf = io.BytesIO() |
| np.savez_compressed(buf, **keep) |
| a.out.write_bytes(buf.getvalue()) |
| print(json.dumps(dict(out=str(a.out), bytes=a.out.stat().st_size, sha256=hashlib.sha256(buf.getvalue()).hexdigest(), |
| arrays={k: list(v.shape) for k, v in keep.items()}, device=nn.info()["gpus"][0]), indent=1)) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
