It's better if you don't want to spend time on infrastructure. The process of running multiple experiments for fine-tuning or distributed experiments on multi-nodes could become a big problem. Also sharing resources with the rest of the team (Memory, CPU, and GPU) is not completely straightforward, and most of the time requires an in house solution. This is why we built polyaxon, to abstract all these engineering work, so that data scientists can focus on developing machine learning and deep learning algorithms, without worrying too much about infrastructure.
