Hacker News Reader

Hacker News

Comment by dweekly | Hacker News Reader

Parent Full thread

dweekly

The relentlessness of the pace of ML remains breathtaking.

Microsoft: We beat human performance and lapped everyone else!

XLNet: Hold my beer.

Code: https://github.com/zihangdai/xlnet

Description: https://towardsdatascience.com/what-is-xlnet-and-why-it-outp...

Paper: https://arxiv.org/pdf/1906.08237.pdf

View on HN

Notably, this result would be very hard for Microsoft to achieve, or indeed even reproduce fully in-house, because it requires more memory than GPUs have. GitHub mentions that TPUs are pretty much table stakes to train this.

I guess they can use TPUs with GCE.

> because it requires more memory than GPUs have

could you elaborate? TPU V3 unit has 16GB of memory, and old V100 also has 16GB of memory. Plus TPU has extra memory consumption for mandatory tensor padding, which GPU doesn't have.

The XLNet GitHub elaborates pretty well, no need to duplicate it here: https://github.com/zihangdai/xlnet

There is no elaboration of your statement there. It says it is hard to reproduce results on single 16GB GPU, but nor they used single TPU in their result. They explicitly state that large number of GPU required: "Therefore, a large number (ranging from 32 to 128, equal to batch_size) of GPUs are required to reproduce many results in the paper.", and they used around 200 TPUs themself according to their publication.

Yes, it boils down to using either a single machine with multiple TPUs (and therefore doing things "the easy way" and relatively quickly) or having to use 128 GPUs (up to 8 per machine) and working with a single sample per GPU, really, really slowly. Given that a single model often requires dozens, if not hundreds of training runs to figure out hyperparameters that achieve a SOTA result, this means that with TPU you can do this, and with GPU you aren't even going to bother training something this big because it will take forever, and someone with a TPU will figure out something better by the time you're done. Which is what happened in this case, it looks like.

The memory constraint is certainly an issue, but I think they can be overcome with some software hacks (for a speed penalty, of course.) For example, gradient checkpointing and gradient accumulation might help:

https://medium.com/tensorflow/fitting-larger-networks-into-m... https://medium.com/huggingface/training-larger-batches-pract...

Reply on news.ycombinator.com