Screening Data Points for Empirical Risk Minimization

This is the code associated to the AISTATS 2020 paper Screening Data Points in Empirical Risk Minimization via Ellipsoidal Regions and Safe Loss Functions.

API

The API is similar to scikit-learn.

from sklearn.model_selection import train_test_split
from utils.loaders import load_experiment

X, y = load_experiment(dataset='cifar10_kernel', synth_params=None, size=10000, redundant=0, 
                        noise=None, classification=True)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)

# random initialization of the ellipsoid and the screener object.
z_init = np.random.rand(X_train.shape[1])
screener = EllipsoidScreener(lmbda=0, mu=0, loss='safe_logistic', penalty='l2', 
                            intercept=False, classification=True, n_ellipsoid_steps=2000, 
                            better_init=20, better_radius=1, cut=False, clip_ell=False, 
                            sgd=False, acceleration=True, dc=False, use_sphere=False,
                            ars=True).fit(X_train, y_train)

prop = np.unique(y_test, return_counts=True)[1]
print('BASELINE : ', 1 - prop[1] / prop[0])
print('SCORE SCREENER : ', screener.score(X_test, y_test))

# screening score of the samples
scores = screener.screen(X_train, y_train)
idx_safeell = np.where(scores > 0)[0]
print('NB TO KEEP', len(idx_safeell))

# we check that the scores are similar for the whole and screened datasets
if len(idx_safeell) !=0:
    estimator_whole = fit_estimator(X_train, y_train, loss='safe_logistic', penalty='l2', 
                                    mu=0, lmbda=0, intercept=False)
    print(y_train[idx_safeell][:10])
    print(estimator_whole.score(X_test, y_test))
    estimator_screened = fit_estimator(X_train[idx_safeell], y_train[idx_safeell], 
                                        loss='safe_logistic', penalty='l2', mu=0, 
                                        lmbda=0, intercept=False)
    print(estimator_screened.score(X_test, y_test))

Installation

- Python 3.6
- numpy
- scipy
- matplotlib
- scikit-learn=0.19
- pandas

Run

export PYTHONPATH=$PWD:$PYTHONPATH

and then

python experiment.py

Possible arguments are detailed in experiment.py.

Interval Regression : Install plasp

git clone https://github.com/VivienCabannes/plasp
cd plasp
pip install -e .