To order to get high device utilization when training, the most important best practice is to both do data prefetching (moving the next batch of data to GPU memory while the previous batch is being processed) and asynchronous logging (moving the metrics from the previous batch to

→ View original post on X — @fchollet,