psmil 0.0.15

psmil package

Partially Subsampled Multiple Instance Learning

This is a NumPy–based toolkit for multiple-instance learning (MIL) with Gaussian mixutre models. This is also a Python package accompanying the following paper:

Yu, B., Li, X., Zhou, J., and Wang, H. Detecting Breast Carcinoma Metastasis on Whole-Slide Images by Partially Subsampled Multiple Instance Learning, The Annals of Applied Statistics, To appear.

PSMIL implements:

1. Instance-supervised MLE (InsMLE) when instance labels are available,
2. Bag-only MLE (BagMLE) when only bag labels are available, and
3. Subsampling MLE (SubMLE) that leverages subsampling to improve statistical efficiency.

For a tutorial and the CAMELYON16 analysis, see the GitHub repo:
https://github.com/Jamesyu420/PSMIL.

Installation

pip install PSMIL

Quick Start

All functions operate on NumPy arrays with shapes:

• X: (N, M, p) float — features for N bags, M instances per bag, and p features
• Y: (N,) int in {0,1} — bag labels
• A: (N, M) int in {0,1} — instance labels

InsMLE

Instance-based MLE when A is observed.

ins = InsMLE(X, A, Y)

Returns a dict with keys:

• 'mu1' (p,) — mean of positives,
• 'mu0' (p,) — mean of negatives,
• 'alpha' float — bag prevalence np.mean(Y),
• 'pi' float — positive-instance rate on positive bags,
• 'Sigma' (p,p) — pooled covariance.

BagMLE

BagMLE when only Y is used.

bag = BagMLE(
    X, Y,
    mu1_init, mu0_init, Sigma_init, pi_init,
    prt=False, iter=100, tol=1e-10
)

Parameters

• mu1_init, mu0_init (p,): initial means
• Sigma_init (p,p): initial covariance (SPD)
• pi_init float in (0,1): initial positive-instance rate on positive bags
• prt bool: print per-iter diagnostics
• iter int: max iterations
• tol float: stopping threshold on parameter change (squared L2 distance sum)

Returns a dict with:

• 'mu1', 'mu0', 'Sigma', 'pi', 'alpha' — estimates
• 'pi_im' (N,M) - Posterior probability
• 'iter' int - iterations

SubMLE

SubMLE to subsample those instances that are likely to be positive.

alpha_n = sub_alphan(X, Y, est=bag, gamma_target=0.05)

sub = SubMLE(
    X, Y, A,
    mu1_init, mu0_init, Sigma_init, pi_init,
    est, alpha_n=alpha_n,
    prt=False, iter=100, tol=1e-10, seed=0
)

Here gamma_target is the target subsampling rate.

Contact

If you have any questions, please feel free to contact Baichen Yu and Prof. Xuetong Li.