boundml-rs (Python)

Small numeric models, packaged with their preprocessing and schema. Train linear, logistic/softmax or small MLP models from CSV, and get one checked package with the feature schema, fitted preprocessing, weights, self-tests and a planned memory budget. Load it and predict from raw values. The computation is native Rust (no PyTorch, no ONNX toolchain), the same code as the BoundML CLI and Rust library, so predictions are bitwise identical across all three.

Early public beta (0.1.0). Dense models on numeric features, CPU only. Project, CLI and Rust library: https://github.com/MahmoudYahiaAhmed/boundml.

Install

pip install boundml-rs

import boundml_rs

The distribution is boundml-rs and the import name is boundml_rs. The project named boundml on PyPI is unrelated.

Use

Training takes a TOML config and CSV files. The CLI release archives include a real example: UCI breast-cancer data, a 30-feature binary classification problem. Download an archive from the v0.1.0 release, extract it, and run this from the extracted folder:

import boundml_rs
import pandas as pd

# Train from a TOML config, then compare with a constant baseline on held-out data.
result = boundml_rs.train("examples/breast_cancer/linear.toml", "bc/ckpt")
print(result["completed"], result["global_step"])                     # True 960
metrics = boundml_rs.evaluate("bc/ckpt", "datasets/breast_cancer/test.csv")
print(metrics["model"]["accuracy"], metrics["baseline"]["accuracy"])  # 0.9596... 0.5806...

# Package it: weights + schema + fitted preprocessing + self-tests, verified after writing.
model = boundml_rs.pack("bc/ckpt", "bc/model.boundml")

# Later, from anywhere: load and predict from raw values.
model = boundml_rs.load("bc/model.boundml")
print(model.verify()["passed"])                       # True
df = pd.read_csv("datasets/breast_cancer/test.csv")   # extra columns (id, diagnosis) are ignored
p = model.predict(df)                                 # float32, shape (124, 1): P(class "M")
print(model.outputs, p.shape, p[:3, 0])               # ['p_M'] (124, 1) [0.9998099  0.99994636 0.9985293 ]

predict accepts a pandas DataFrame (columns matched by name), one dict or a list of dicts (keys are feature names), or a NumPy array of shape (rows, n_features) in model.features order. None or NaN means missing and follows the schema's missing-value policy. The result is a float32 array of shape (rows, len(model.outputs)).

Errors raise boundml_rs.BoundMLError with a stable code, for example BML-D004 (schema mismatch), BML-D005 (a value that is not a number) or BML-C005 (memory budget exceeded). Every code is explained in docs/errors.md.

A Model holds one of the package's declared sessions once it has predicted. It is thread-safe, and predictions release the GIL.

Limits

Dense models only (linear, logistic/softmax, MLPs with ReLU/Tanh/Sigmoid) on numeric features; no categorical features, trees, images, text or GPU. Single-threaded kernels: ONNX Runtime is about 2–2.4× faster on large batches. On the two bundled datasets, the tested MLPs did not beat the linear baselines. Measurements and their limits: docs/benchmarks.md.

License: MIT OR Apache-2.0.