Authors of the paper: Durga Sivasubramanian (durgas@cse.iitb.ac.in), Rishabh Iyer (rishabh.iyer@utdallas.edu), Ganesh Ramakrishnan (ganesh@cse.iitb.ac.in), Abir De (abir@cse.iitb.ac.in)
The code was cleaned and modularized by Sanidhya Anand
This directory contains code necessary to run the SELCON algorithm which is a data subset selection algorithm for efficient training. In a nutshell, it aims to select the training subset as well as the model parameters subject to a set of constraints ensuring that the error on validation set remains below an acceptable level. We show that solving the above problem is equivalent to minimizing a monotone and approximate submodular function.
If you use this code in your paper, please use:
@inproceedings{durga2021training,
title={Training Data Subset Selection for Regression with Controlled Generalization Error},
author={Durga, S and Iyer, Rishabh and Ramakrishnan, Ganesh and De, Abir},
booktitle={International Conference on Machine Learning},
pages={9202--9212},
year={2021},
}
You can install the package using
pip install selcon
To run this code fully, you'll need PyTorch (we're using version 1.4.0) and scikit-learn. We've been running our code in Python 3.7.
SELCON package can be utilised in Linear Subset Selection or Deep Subset Selection methods as:
from SELCON.datasets import load_def_data, get_data
from SELCON.linear import Regression
load_def_data provides functionality for using the datasets used for the experiments in the paper (provided you have them available in the 'Dataset' directory)
reg = Regression()
# Converts specified numpy arrays to torch tensors (assuming data has been split previously)
X_trn, X_val, Y_trn, Y_val = get_data(x_train, x_val, y_train, y_val)
# Trains SELCON model for a subset fraction of 0.03 on the training subset (no fairness)
reg.train_model(X_trn, Y_trn, X_val, Y_val, fraction = 0.03)
# Return optimal subset indices
subset_idxs = reg.return_subset()
# Returns the optimal subset of the training data for further use
X_sub = X_trn[subset_idxs]
y_sub = Y_trn[subset_idxs]
from SELCON.datasets import load_def_data, get_data
from SELCON.deep import DeepSelection
reg = DeepSelection()
# Converts specified numpy arrays to torch tensors (assuming data has been split into train-val sets previously)
X_trn, X_val, Y_trn, Y_val = get_data(x_train, x_val, y_train, y_val)
# Trains SELCON model for a subset fraction of 0.03 on the training subset (with fairness)
reg.train_model_fair(X_trn, Y_trn, X_val, Y_val, fraction = 0.03)
# Return optimal subset indices
subset_idxs = reg.return_subset()
# Returns the optimal subset of the training data for further use
X_sub = X_trn[subset_idxs]
y_sub = Y_trn[subset_idxs]