Paper: Finding High-Value Training Data Subset through Differentiable Convex Programming
Authors: Soumi Das, Arshdeep Singh, Saptarshi Chatterjee, Suparna Bhattacharya, Sourangshu Bhattacharya
This is the code repository for the paper "Finding High-Value Training Data Subset through Differentiable Convex Programming" accepted at ECML-PKDD 2021 (https://2021.ecmlpkdd.org/). We propose the method HOST-CP (High-value Online Subset selection of Training samples through differentiable Convex Programming) for selecting subsets in an online method.
- Python, NumPy, PyTorch, cvxpy, cvxpylayers, faiss, Pillow, scikit-learn, Matplotlib
This setup is provided for CIFAR10 dataset using ResNet-18 model.
One needs to first store ResNet-18 checkpoint under resnet_checkp/ which is pre-trained on the entire dataset. python train.py can be used for this purpose.
Following this, running python subsetfind.py yields the subset based on the fraction one provides.
python trainsub.py can be used to run training on the subset provided by the method.