Pytorch code for simultaneous classifier learning and sparse coding.
This package includes the code used for the following experiments in the paper "Learning from Incomplete Features by Simultaneous Training of Neural Networks and Sparse Coding" by CF Caiafa, Z Wang, J Solé-Casals and Q Zhao, accepted for presentation at L2ID Workshop at CVPR 2021 (19 - 25 June, 2021). Preprint available at: https://arxiv.org/abs/2011.14047
- Synthetic dataset with a logistic regression classifier (1-layer NN) We compare our simultaneous method with standard imputation methods:
Seq. Sparse: reconstructions are obtained by finding the sparsest representation compatible with the observations solving a LASSO problem;
Zero Fill: missing entries are filled with zeros, which is equivalent to ignore unknown values;
Mean Unsupervised: missing entries are filled with the mean computed on the available values in the same position in the rest of data samples;
Mean Supervised: as in the previous case but the mean is computed on the samples of the same class vectors only;
K-Nearest Neighbor (KNN): as in the previous case but the mean is computed on the K-Nearest Neighbors of the same class only
-
MNIST dataset with a CNN4 classifier (adapted from https://github.com/pytorch/examples/tree/master/mnist)
-
CIFAR10 dataset with a ResNet18 classifier (adapted from https://github.com/kuangliu/pytorch-cifar)
Dependencies:
- Pytorch 1.0.0
