sawmil 0.1.5

Sparse Multiple-Instance Learning: SVM, NSK, sMIL and sAwMIL.

Sparse Multiple-Instance Learning in Python

[!Warning]

It is an alpha version of the package.

MIL models based on the Support Vector Machines (NSK, sMIL, sAwMIL). Inspired by the outdated misvm package.

Note: This is an alpha version.

Implemented Models

Normalized Set Kernels (NSK)

Gärtner, Thomas, Peter A. Flach, Adam Kowalczyk, and Alex J. Smola. Multi-instance kernels. Proceedings of the 19th International Conference on Machine Learning (2002).

Sparse MIL (sMIL)

Bunescu, Razvan C., and Raymond J. Mooney. Multiple instance learning for sparse positive bags. Proceedings of the 24th International Conference on Machine Learning (2007).

Sparse Aware MIL (sAwMIL)

Classifier used in trilemma-of-truth:

Savcisens, Germans, and Tina Eliassi-Rad. The Trilemma of Truth in Large Language Models. arXiv preprint arXiv:2506.23921 (2025).

Here’s a tighter, clearer install section you can drop into your README.

---

Installation

sawmil supports two QP backends: Gurobi and OSQP. By default, the base package installs without any solver; pick one (or both) via extras.

Base package (no solver)

pip install sawmil

Option 1 — Gurobi backend

Gurobi is commercial software. You’ll need a valid license (academic or commercial), refer to the official website.

pip install "sawmil[gurobi]"
# it installs numpy>=1.22 and scikit-learn>=1.7.0

Option 2 — OSQP backend

pip install "sawmil[osqp]"
# in additionl to the base packages, it installs osqp>=1.0.4 and scipy

Option 3 — All supported solvers

pip install "sawmil[full]"

Picking the solver in code

from sawmil import SVM

# solver= "osqp" (default is "gurobi")
clf = SVM(C=1.0, kernel="rbf", gamma=0.5, solver="osqp").fit(X, y)

Requirements

numpy>=1.22
scikit-learn>=1.7.0

Quick start

1. Generate dummy data

from dataset import make_complex_bags
import numpy as np
rng = np.random.default_rng(0)

ds = make_complex_bags(
    n_pos=300, n_neg=100, inst_per_bag=(5, 15), d=2,
    pos_centers=((+2,+1), (+4,+3)),
    neg_centers=((-1.5,-1.0), (-3.0,+0.5)),
    pos_scales=((2.0, 0.6), (1.2, 0.8)),
    neg_scales=((1.5, 0.5), (2.5, 0.9)),
    pos_intra_rate=(0.25, 0.85),
    ensure_pos_in_every_pos_bag=True,
    neg_pos_noise_rate=(0.00, 0.05),
    pos_neg_noise_rate=(0.00, 0.20),
    outlier_rate=0.1,
    outlier_scale=8.0,
    random_state=42,
)

2. NSK with RBF Kernel

Load a kernel:

from sawmil.kernels import get_kernel
from sawmil.bag_kernels import make_bag_kernel
k = get_kernel("rbf", gamma=0.5) # base (single-instance kernel)
bag_k  = make_bag_kernel(k, use_intra_labels=False) # convert single-instance kernel to bagged kernel

Fit NSK Model:

from sawmil.nsk import NSK

clf = NSK(C=0.1, bag_kernel=bag_k, scale_C=True, tol=1e-8, verbose=False).fit(ds, None)
print("Train acc:", clf.score(ds, np.array([b.y for b in ds.bags])))

3. Fit sMIL Model with Linear Kernel

from src.sawmil.smil import sMIL

k = get_kernel("linear", normalizer="none") # base (single-instance kernel)
bag_k  = make_bag_kernel(Linear(), normalizer="none", use_intra_labels=False)
clf = sMIL(C=0.1, bag_kernel=bag_k, scale_C=True, tol=1e-6, verbose=False).fit(ds, None)

print("Train acc:", clf.score(ds, np.array([1 if b.y > 0 else -1 for b in ds.bags])))

See more examples in the example.ipynb notebook.

Citation

If you use sawmil package in academic work, please cite:

Savcisens, G. & Eliassi-Rad, T. sAwMIL: Python package for Sparse Multiple-Instance Learning (2025).

@software{savcisens2025sawmil,
  author = {Savcisens, Germans and Eliassi-Rad, Tina},
  title = {sAwMIL: Python package for Sparse Multiple-Instance Learning},
  year = {2025},
  doi = {10.5281/zenodo.16990499},
  url = {https://github.com/carlomarxdk/sawmil}
}

If you want to reference a specific version of the package, find the correct DOI here.