Step 13 of 16Day 21 · Sun 15 NovWeek 3Phase A · notebooks

Adoption gate

The go/no-go check before any web code. All four model types must run end to end from config files, from a clean kernel, with numbers you can repeat.

Runs on: JupyterLab on the Jetson
Notebooks: 90_runner.ipynb (final)
Creates: configs/detect.yaml, reports/gate.md, git tag phase-a-done
Time: Day 21

Where Where to implement

Runner

Extend sutra_core/run.py so type: image and type: object_detection call your Step 04 and Step 09 code

Report

reports/gate.md: one row per model, generated by the notebook

Do Do it

1Add the detection config

Fileconfigs/detect.yaml

name: detect_v1
dataset: data/raw/detect
input:      { type: image, width: 96, height: 96, color: grayscale }
processing: { type: image, normalize: true }
learning:   { type: object_detection, arch: centroid_mobilenetv2_050, epochs: 60, lr: 0.001, object_weight: 20 }
export:     [onnx, tflite_int8, tensorrt]

2Route image and detection through run()

Wrap your Step 04 training code in train_image(cfg, out) and the Step 09 code in train_detect(cfg, out), both in sutra_core/train.py. Each returns a metrics dict. Then dispatch on the input or learning type:

Filesutra_core/run.py (inside run(), first lines)

cfg = yaml.safe_load(open(config_path))
out = Path(out_dir) if out_dir else ROOT
if cfg["learning"]["type"] == "object_detection":
    from .train import train_detect
    return train_detect(cfg, out)
if cfg["input"]["type"] == "image":
    from .train import train_image
    return train_image(cfg, out)
# ... timeseries path from Step 08 continues here

3Run all four from a clean kernel

Notebook cellnotebooks/90_runner.ipynb

import sys, json, time
from pathlib import Path
ROOT = Path.home() / "sutraai"; sys.path.insert(0, str(ROOT))
from sutra_core.run import run

results, timing = {}, {}
for name in ["image", "sensor", "audio", "detect"]:
    t = time.perf_counter()
    results[name] = run(ROOT / "configs" / f"{name}.yaml")
    timing[name] = round((time.perf_counter() - t) / 60, 1)

score = lambda m: m.get("test_f1", m.get("test_acc"))
lines = ["| model | metric | test score | minutes |", "|---|---|---|---|"]
for k, m in results.items():
    lines.append(f"| {k} | {'F1' if 'test_f1' in m else 'accuracy'} | {score(m):.3f} | {timing[k]} |")
(ROOT / "reports/gate.md").write_text("\n".join(lines)); print("\n".join(lines))

4Repeat once to prove it's stable

Kernel → Restart and Run All a second time. Every score should land within about 1 point (0.01 F1) of the first run. If not, check your seeds and data splits.

5Decide

All boxes below ticked: tag the repo and start Step 14 tomorrow.

Jetson terminal

cd ~/sutraai && git add -A && git commit -m "Phase A complete" && git tag phase-a-done

Something failing: use Days 22–23 to fix it and slide Phase B by up to two days. Don't start FastAPI with a broken model: debugging ML through a web stack is much harder.

Test Test it

Notebook cellnotebooks/90_runner.ipynb (last cell)

assert results["image"]["test_acc"] >= 0.85
assert results["sensor"]["test_acc"] >= 0.85
assert results["audio"]["test_acc"] >= 0.80
assert results["detect"]["test_f1"] >= 0.75
print("✅ ADOPTION GATE PASSED: go to Phase B")

Done The gate

NB00 is green (GPU, cuDNN, TensorRT, ONNX Runtime CUDA)
Image, sensor, audio and detection each go train → evaluate → export → int8 check
Every int8 model: accuracy drop under 3 points (detection: F1 drop under 0.05)
All four run from configs/*.yaml via run(), from a clean kernel
Second run reproduces scores within about 1 point
MFCC golden vectors and profile.md saved
Device simulator: v1 → v2 update, sha256 refusal, telemetry all work
Nemotron advisor correct; RAG hit@4 ≥ 8/10
Reusable code lives in sutra_core/, not only in notebook cells
BUILDLOG has the numbers; repo tagged phase-a-done

If it breaks

Symptom	Fix
One model passes in its own notebook but fails in run()	A setting differs between the notebook and the YAML (size, lr, epochs, normalization). Diff them.
Scores jump between runs	Seed `random`, `numpy` and `torch` at the top of run(); keep splits by file and seeded.
Kernel dies mid-run	Free GPU memory between models: `del model; torch.cuda.empty_cache()` at the end of each train function.

Learn more

Section 10 of the beginner guide: the same checklist with context