1fabd4a246
The model layer of the project, built honestly:
- tools/dataset_validate.py — full-sweep validator over the receiver
store (sha256, schema, monotonic labels, telemetry-row gate). On the
current corpus: 64,798 accepted + 8,154 degraded + 3,701 rejected +
7 errored across 76,660 shipped episodes. data/processed/validation_v1.parquet
is committed as the per-episode acceptance index.
- training/_features.py — channel registry (46 channels across
proc/guest/qmp/netflow), summary-stat windowing AND channel×time
tensor extraction at 10s/5s windowing. Time alignment uses t_wall_ns
(Unix ns) — tested fix for a real netflow-vs-host clock-base
inconsistency that was silently dropping every netflow channel.
- training/_split.py — three held-out recipes (host / sample / time)
with profile-stratification assertions. held_out_host carries
untested_profiles for cases like scan-and-dial absent from the test
host (5 of 6 profiles tested cross-device, never silently averaged).
- training/models/ — 6 architectures behind a common BaseModel
interface: gbt (XGBoost), mlp, cnn, gru, lstm, transformer. Each
trained twice (realistic / oracle) per the deployment threat model.
Schema-hashed checkpoints refuse to load if _features.py changed
since training (silent-input-drift protection, tested).
- training/trainer/ — unified training loop: class-weighted CE, LR
warmup + cosine, gradient clipping, mixed precision when CUDA,
early stopping on val macro F1, best-on-val checkpoint. Same loop
runs MLP/CNN/GRU/LSTM/Transformer; GBT uses XGBoost
early_stopping_rounds on val mlogloss.
- training/eval_/ — bootstrap 95% CIs on macro F1, per-class F1,
per-profile and per-host breakdown, paired-bootstrap significance
for model-vs-model gap. Confusion matrix uses union of seen labels.
- training/dashboard/producers/ — replay/metrics/perf/profiles
emitting the six event types the dashboard's awaiting scenes
consume; on-demand tensor extraction so the Pi can run live
inference without 65 GB of shards.
- 17 unit tests (split coverage, features round-trip, schema mismatch,
determinism, time-base alignment regression).
End-to-end smoke-trained all six on a 567-episode subset; held-out
test macro F1 reported with paired-bootstrap significance. The
methodology now reports honest cross-device generalization, not
in-distribution validation.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
46 lines
1.4 KiB
Python
46 lines
1.4 KiB
Python
"""Model registry — name → builder.
|
|
|
|
Importing the architecture modules has side effects (registers each
|
|
class with REGISTRY) so callers can do::
|
|
|
|
from training.models import get_model
|
|
cls = get_model("cnn")
|
|
|
|
without knowing which file defines it.
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
from typing import Any, Callable
|
|
|
|
REGISTRY: dict[str, Callable[..., "BaseModel"]] = {}
|
|
|
|
|
|
def register(name: str):
|
|
def decorator(cls):
|
|
if name in REGISTRY:
|
|
raise ValueError(f"model {name!r} already registered")
|
|
REGISTRY[name] = cls
|
|
cls.__model_name__ = name
|
|
return cls
|
|
return decorator
|
|
|
|
|
|
def get_model(name: str):
|
|
if name not in REGISTRY:
|
|
raise KeyError(
|
|
f"model {name!r} not registered; known: {sorted(REGISTRY)}"
|
|
)
|
|
return REGISTRY[name]
|
|
|
|
|
|
# Eager-import the implementations so the registry is populated.
|
|
# Order matters only for which "kind" gets imported first — all are listed.
|
|
from training.models import gbt # noqa: F401,E402
|
|
from training.models import mlp # noqa: F401,E402
|
|
from training.models import cnn # noqa: F401,E402
|
|
from training.models import gru # noqa: F401,E402
|
|
from training.models import lstm # noqa: F401,E402
|
|
from training.models import transformer # noqa: F401,E402
|
|
from training.models import transformer_ssl # noqa: F401,E402
|
|
|
|
from training.models._base import BaseModel # noqa: E402,F401
|