Gradient-based learning for the $F$-measure and other performance metrics

System requirements

We only tested our implementation on Ubuntu 16.04 and Python 3.6.5. Please follow website instructions to install pytorch 1.0.0 and torchvision 0.2.1 (GPU recommended). Please install the following packages via pip:

pip install -U tensorflow
pip install tensorboardX
pip install -U protobuf

Datasets

Please download the Adult, CIFAR10, CIFAR100, Covertype, KDDCup08, Letter and MNIST dataset:

bash data.sh

Experiments

Binary and multi-class classification experiments are implemented respectively in binary.py and multi.py. To reproduce results in our paper:

Binary classification, F1 score, and linear classifier

bash binary.sh f1 linear NUMBER-OF-PROCESSES
for x in f1-linear-*; do bash select.sh $x f1; done

Binary classification, F1 score, and multi-layer perceptron

bash binary.sh f1 mlp NUMBER-OF-PROCESSES
for x in f1-mlp-*; do bash select.sh $x f1; done

Multi-class classification, micro F1 score, and linear classifier

bash multi.sh f1_micro linear NUMBER-OF-PROCESSES
for x in f1_micro-linear-*; do bash select.sh $x f1; done

Multi-class classification, micro F1 score, and multi-layer perceptron

bash multi.sh f1_micro mlp NUMBER-OF-PROCESSES
for x in f1_micro-mlp-*; do bash select.sh $x f1; done

Binary classification, G-measure (a.k.a. Fowlkes–Mallows index), and linear classifier

bash binary.sh g1 linear NUMBER-OF-PROCESSES
for x in g1-linear-*; do bash select.sh $x g1; done

Binary classification, G-measure (a.k.a. Fowlkes–Mallows index), and multi-layer perceptron

bash binary.sh g1 mlp NUMBER-OF-PROCESSES
for x in g1-mlp-*; do bash select.sh $x g1; done

Multi-class classification, G-measure (a.k.a. Fowlkes–Mallows index), and linear classifier

bash multi.sh g1_micro linear NUMBER-OF-PROCESSES
for x in g1_micro-linear-*; do bash select.sh $x g1; done

Multi-class classification, G-measure (a.k.a. Fowlkes–Mallows index), and multi-layer perceptron

bash multi.sh g1_micro mlp NUMBER-OF-PROCESSES
for x in g1_micro-mlp-*; do bash select.sh $x g1; done