The resource restrictions for this challenge are designed to reflect real-world conditions where machine learning models must operate efficiently on low-resource devices, such as smartphones.

The T4 GPU restriction ensures a fair playing field, avoiding disparities between participants with access to different levels of hardware. While it may seem like restricting only inference resources would be sufficient, setting a uniform limit on training hardware also ensures that all participants develop models under similar constraints, pushing for optimization in both training and inference phases. Furthermore, the challenge’s focus on edge devices requires models that are resource-efficient at all stages, from development to deployment.

If my calculations are correct, 3 hours inference is equivalent to ~1 second per label. If an image has 20 labels that is ~20 seconds for an image which is an acceptable time. Of course, if you can optimise that further that would be amazing! However the 9 hours training time remains.