Another version of the algorithm can be "decentralized" where the resulting model weights get averaged:
Those processes are synchronous because we need to wait for all the workers to finish their jobs. The same processes can happen asynchronously, only the gradients or weights are not averaged. You can learn more about it here: https://arxiv.org/pdf/2007.03970.pdf When it comes to the centralized synchronous approach, Pytorch and TensorFlow seem to follow a slightly different strategy (https://pytorch.org/docs/stable/notes/ddp.html) as it doesn't seem to be using a parameter server as the gradients are synchronized and averaged on the worker processes. This is how the Pytorch DistributedDataParallel module is implemented (https://pytorch.org/.../torch.nn.parallel...), as well as the TensorFlow MultiWorkerMirroredStrategy one (https://www.tensorflow.org/.../MultiWorkerMirroredStrategy). It is impressive how simple they made training a model in a distributed fashion!
此处可能存在不合适展示的内容,页面不予展示。您可通过相关编辑功能自查并修改。
如您确认内容无涉及 不当用语 / 纯广告导流 / 暴力 / 低俗色情 / 侵权 / 盗版 / 虚假 / 无价值内容或违法国家有关法律法规的内容,可点击提交进行申诉,我们将尽快为您处理。