Book contents +

XVII · Python C-ABI (welvet on PyPI) / CHAPTER 28

Python C-ABI: train SGD and .entity round-trip

Samuel Watson · Welvet feature bookOriginal source ↗

welvet (PyPI)cpu_tiled


Why it exists

Training and checkpoints should match the Go ENTITY path so Python scripts can hand off to Go apps (and back).

What it is

Train one Dense step with train_sgd, serialize to .entity bytes, deserialize a second grid, and free both handles.

Checkpoint bytes

serialize_entity returns the same Welvet .entity blob the Go engine writes. deserialize_entity returns a new grid handle: always free_grid both. For cameral / named train modes see Go book §67 and Python list_named_train_modes() / train_stack_mse.

Requirements: chapter 114.

Python example

examples/116-python-train-entity/main.py

Run:cd welvet/examples/116-python-train-entity && python3 main.py
"""116. Python C-ABI: train SGD and .entity round-trip."""
from __future__ import annotations

import welvet as w


def main() -> None:
    w.assert_engine_version()
    g = w.create_grid(backend="cpu_tiled")
    try:
        w.place_dense(
            g,
            {"in": 8, "out": 4, "act": "relu", "dtype": w.DType.FLOAT32},
        )
        x = [0.05 * (i + 1) for i in range(8)]
        y = [0.0, 0.0, 0.0, 1.0]
        result = w.train_sgd(g, x, y, lr=0.05)
        print("train", result)

        blob = w.serialize_entity(g)
        print("entity_bytes", len(blob))
    finally:
        w.free_grid(g)

    g2 = w.deserialize_entity(blob)
    try:
        out = w.forward(g2, x)
        print("reloaded_out", [round(v, 4) for v in out])
    finally:
        w.free_grid(g2)


if __name__ == "__main__":
    main()

Output

exit 0 · last run via python3 main.py

train {'loss': 0.25, 'status': 'ok'}
entity_bytes 550
reloaded_out [0.0, 0.0, 0.0, 0.0127]

Validate (harness)

cd welvet/examples/116-python-train-entity && PYTHONPATH=../../../w2a/python/src python3 main.py