# How to maximize subset accuracy for multilabel multiclass image classification

I working on multi-label multi-class image classification. I am using TensorFlow. Currently I am using sigmoid on output layer with binary_crossentrpy. Model is giving fair accuracy, but for my requirement I need to rely on Subset accuracy (exact match) and not overall accuracy. For evaluation purpose I am using Subset Accuracy metric, but model is obviously maximizing overall accuracy. How can I add a custom loss function so that I can maximize subset accuracy.

• You mean for a specific subset or the subset accuracy for all subsets? – Janosch Jan 17 at 9:42
• @Janosch Look for Exact Match aka subset accuracy. Currently I am calculating subset accuracy for my test set. – Vivek Mehta Jan 17 at 9:55