A PyTorch implementation of "LOW: Training Deep Neural Networks by Learning Optimal Sample Weights" with an example on MNIST using LeNet-5.

1. Original submission

• PyTorch
• cvxopt
• numpy

Running the demo:

• LOW on MNIST:

python main.py --data <path_to_folder_with_mnist> --save <path_to_save_dir> --loss 'LOW'

• Standard CE on MNIST:

python main.py --data <path_to_folder_with_mnist> --save <path_to_save_dir> --loss 'CE'

Options:

• --loss (str) - available losses: 'CE' and 'LOW' (default)
• --n_epochs (int) - number of epochs (default 100)
• --batch_size (int) - size of minibatch (default 256)

A comparison between SGD and LOW:

@article{unpublished,
  title={LOW: Training Deep Neural Networks by Learning Optimal Sample Weights},
  author={Santiago, C. and Barata, C. and Sasdelli, M. and Carneiro, G. and Nascimento, J.},
  journal={Pattern Recognition},
  pages={1--1},
  year={2020},
  publisher={Elsevier}
}