SutraCore/Learn/SutraAI Jetson Lab
Free course · by the SutraCore team

Free NVIDIA Jetson AI course: 30 days, 16 steps

Orin NX 16GBJetPack 6.2.1CUDA 12.6TensorRT 10.3Python 3.10PyTorch 2.8
Step 14 of 16Day 22 · Mon 16 NovWeek 4Phase B · web

Notebooks to an ml_core package

Turn sutra_core into an installable package with a command-line entry point and tests. FastAPI will call this package: no notebook code gets copied into the web app.

Runs on
Jetson terminal (or VS Code Remote-SSH)
Notebooks
None: plain Python from here on
Creates
packages/ml_core/, sutra-run command, tests/
Time
Day 22

Where Where to implement

Package

  • ~/sutraai/packages/ml_core/src/ml_core/
  • Move (don't rewrite) the files from sutra_core/

Tests

  • packages/ml_core/tests/
  • Golden MFCC vectors + a tiny smoke dataset

Do Do it

1Create the layout and move the code

Jetson terminal
cd ~/sutraai
mkdir -p packages/ml_core/src/ml_core packages/ml_core/tests
git mv sutra_core/*.py packages/ml_core/src/ml_core/
ls packages/ml_core/src/ml_core/     # __init__.py data.py dsp.py models.py train.py export.py profile.py device_sim.py detect.py run.py
Filepackages/ml_core/pyproject.toml
[project]
name = "ml-core"
version = "0.1.0"
description = "SutraAI pipeline engine: input → DSP → learning → export"
requires-python = ">=3.10,<3.11"          # matches the cp310 Jetson wheels on JetPack 6.2.1
# torch 2.8.0 / torchvision 0.23.0 / onnxruntime-gpu 1.23.0 are NOT listed: they come from the Jetson
# jp6/cu126 index (Step 02) and stay pinned by constraints-jp621.txt (PIP_CONSTRAINT in the venv)
dependencies = ["numpy==1.26.4", "pyyaml", "timm", "onnx", "scikit-learn", "librosa", "pandas", "thop", "ai-edge-litert"]

[project.scripts]
sutra-run = "ml_core.run:main"

[build-system]
requires = ["hatchling"]
build-backend = "hatchling.build"

[tool.hatch.build.targets.wheel]
packages = ["src/ml_core"]

2A proper command-line entry point

The web worker will run this command and read its JSON lines, the same lines you printed in every training loop.

Filepackages/ml_core/src/ml_core/run.py (replace the old if __name__ block)
import argparse

def main():
    p = argparse.ArgumentParser(prog="sutra-run", description="Train + export one SutraAI pipeline")
    p.add_argument("config", help="path to pipeline.yaml")
    p.add_argument("--data", help="dataset folder (overrides 'dataset' in the YAML)")
    p.add_argument("--out", help="folder for models/, reports/, exports/", default=None)
    a = p.parse_args()
    cfg_path = a.config
    if a.data:                                   # write a copy of the config pointing at the uploaded data
        cfg = yaml.safe_load(open(a.config)); cfg["dataset"] = str(Path(a.data).resolve())
        cfg_path = Path(a.out or ".") / "pipeline.resolved.yaml"
        Path(cfg_path).parent.mkdir(parents=True, exist_ok=True); yaml.safe_dump(cfg, open(cfg_path, "w"))
    run(cfg_path, a.out)

if __name__ == "__main__":
    main()

In load_timeseries, use Path(cfg["dataset"]) directly when it's an absolute path, and only prefix ROOT when it's relative: root = Path(cfg["dataset"]); root = root if root.is_absolute() else ROOT / root.

3Install in editable mode and update imports

Jetson terminal
cd ~/sutraai && source .venv/bin/activate
pip install -e packages/ml_core
grep -rl "sutra_core" notebooks packages | xargs sed -i 's/sutra_core/ml_core/g'
sutra-run --help

4Tests that protect what you proved

Filepackages/ml_core/tests/test_dsp_golden.py
import glob, os, numpy as np
from ml_core.dsp import mfcc

GOLD = os.path.expanduser("~/sutraai/exports/golden")

def test_mfcc_matches_golden_vectors():
    files = glob.glob(f"{GOLD}/mfcc_*.npz")
    assert files, "no golden vectors found: run notebook 31 first"
    for f in files:
        d = np.load(f)
        np.testing.assert_allclose(mfcc(d["x"]), d["y"], rtol=1e-5, atol=1e-5)
Filepackages/ml_core/tests/test_cli_smoke.py
import json, subprocess, sys, os

def test_audio_pipeline_runs_end_to_end(tmp_path):
    cfg = os.path.expanduser("~/sutraai/configs/audio.yaml")
    out = subprocess.run([sys.executable, "-m", "ml_core.run", cfg, "--out", str(tmp_path)],
                         capture_output=True, text=True, timeout=3600)
    assert out.returncode == 0, out.stderr[-2000:]
    done = [json.loads(l) for l in out.stdout.splitlines() if l.startswith("{") and '"event": "done"' in l]
    assert done and done[-1]["test_acc"] > 0.5
Jetson terminal
pip install pytest && cd ~/sutraai && pytest packages/ml_core/tests -q

Test Test it

Jetson terminal
cd ~/sutraai
sutra-run configs/audio.yaml --out /tmp/sutra_try | tail -n 3      # last line: {"event": "done", ...}
pytest packages/ml_core/tests -q                                     # all green
Pass: sutra-run reproduces the NB90 audio score (± about 1 point), prints one JSON line per epoch plus a final "done" line, and the tests pass.

Done Done when

  • packages/ml_core installed with pip install -e
  • Notebooks import ml_core and still pass Restart and Run All
  • sutra-run CONFIG --data DIR --out DIR works
  • Golden-vector test and CLI smoke test pass
  • Committed

If it breaks

SymptomFix
pip tries to install a CPU torchtorch must not be in dependencies, and echo $PIP_CONSTRAINT must print the constraints-jp621.txt path (reactivate the venv if it's empty).
ModuleNotFoundError: ml_core in a notebookRestart the kernel after pip install -e; remove old sys.path.insert lines.
Golden test failsSomeone changed mfcc(). That's exactly what the test is for: revert, or regenerate the vectors on purpose and note why.

Learn more