SRAAL is an open-source code for an adversarial active learning algorithm based on state relabeling. The algorithm describes active learning in the case of limited data labeling budget. The generative model VAE is used to carry out unsupervised feature reconstruction learning based on the variational process on the data, and the state relabeling method is used to calibrate the uncertainty score of the unlabeled samples, so that Using the discriminator to use the confrontation mechanism to evaluate the sampling value of unlabeled samples, so as to select more instructive samples for labeling, which significantly improves the sampling quality and target performance of active learning. The code implements common data sets CIFAR-10, CIFAR-100 and large-scale data sets…

#Adversarial #active #learning #data #annotation #sampling #model #SRAAL #based #state #relabeling

Leave a Comment

Your email address will not be published. Required fields are marked *