Re:Evaluation of Deep Learning Toolkits
by zer0nes on Tuesday December 29, 2015 @06:28PM
(#10809762)
I'm a fan of Keras too. Including comments on Keras and other higher-level libraries would make my original review too long, hence I dropped it.
Note that the performance review is very much incomplete. As mentioned in the blog: Deep Learning is not just about feed-forward convnets, not just about ImageNet, and certainly not just about a few passes over the network. However, Soumith’s benchmark is the only notable one as of today. So we will base the Single-GPU performance rating based on his benchmark.
For TF, I think a bigger single-node perf issue is memory allocation. At NIPS, Jeff Dean didn't have a straight forward answer for why TF's memory perf is so poor.
Note that the performance review is very much incomplete. As mentioned in the blog: Deep Learning is not just about feed-forward convnets, not just about ImageNet, and certainly not just about a few passes over the network. However, Soumith’s benchmark is the only notable one as of today. So we will base the Single-GPU performance rating based on his benchmark.
For TF, I think a bigger single-node perf issue is memory allocation. At NIPS, Jeff Dean didn't have a straight forward answer for why TF's memory perf is so poor.