This repository provides the implementation for "Pointwise Binary Classification with Pairwise Confidence Comparisons" (submitted to ICLR 2021).


Requirements:
Python 3.6.10
numpy 1.14
Pytorch 1.4.0
tologistichvision 0.2

DEMO

-ds: specifies the dataset
-mo: specifies the model
-me: specifies the method
-lo: specifies the loss function (default: logistic loss)
-lr: learning rate
-wd: weight decay
-uci: whether using uci datasets
-gpu: specifies gpu id
-prior: specifies class prior

For example:

python main.py -lo logistic -me PcompTeacher -uci 0 -prior 0.5 -mo resnet -ds cifar10 -lr 1e-3 -wd 1e-3 -gpu 0

python main.py -lo logistic -me PcompUnbiased -uci 1 -prior 0.5 -mo linear -ds pendigits -lr 1e-3 -wd 1e-3 -gpu 0





Thank you!
