google-research/task_adaptation

This repository contains code for evaluating visual models on a challenging set of downstream vision tasks, coming from diverse domains: natural images, artificial environments (structured) and images captured with non-standard cameras (specialized). These tasks, together with our evaluation protocol, constitute VTAB, short for Visual Task Adaptation Benchmark. Our benchmark expects a pretrained model as an input. The model should be provided as a Hub module. The given model is independently fine tuned for solving each of the above 20 tasks.