Build & Run#

Local install#

Requires Python 3.10+ and pip >= 23.

git clone <repo>
cd kanx
pip install -e .[dev,api]    # editable install + dev + api extras

Smoke-test:

python -c "from kanx import KAN; import tensorflow as tf; print(KAN([2,8,1])(tf.zeros((1,2))))"

Run the test suite#

bash scripts/test.sh
# or, equivalently:
pytest tests/ -v --cov=src/kanx

Expected: 113 passed, ~94% coverage on the library.

Train from the CLI#

bash scripts/train.sh                    # uses configs/default.yaml
bash scripts/train.sh configs/mnist.yaml # custom config

The checkpoint is written to checkpoints/kanx_model.keras by default (configurable in YAML).

Benchmark#

bash scripts/benchmark.sh                # regenerates benchmarks/results.md

Serve the REST API locally#

uvicorn api.app:app --host 0.0.0.0 --port 8000 --reload
curl http://localhost:8000/api/health

Docker#

docker build -t kanx:latest .
docker run --rm -p 8000:8000 \
    -e KANX_CHECKPOINT=/app/checkpoints/kanx_model.keras \
    -v $(pwd)/checkpoints:/app/checkpoints \
    kanx:latest

Or via compose:

docker compose up --build

Kubernetes#

docker build -t kanx:latest .
kind load docker-image kanx:latest    # if you use kind
kubectl apply -f k8s/
kubectl port-forward svc/kanx-api 8000:80
curl http://localhost:8000/api/health

See k8s/ — Deployment with rolling updates, Service, Ingress, HPA, PVC for model storage.

Environment variables#

Variable	Default	Used by
`KANX_CONFIG`	`configs/default.yaml`	`api/app.py` startup
`KANX_CHECKPOINT`	`checkpoints/kanx_model.keras`	`api/app.py` startup
`KANX_MAX_BATCH`	`4096`	`/api/predict` guard
`TF_CPP_MIN_LOG_LEVEL`	`2` (recommended in prod)	TensorFlow stderr noise

Reproducibility#

from kanx import set_global_seed
set_global_seed(42)

Seeds Python random, numpy, tf.random and tf.keras.utils.set_random_seed in one call. Called automatically by kanx.train.