Jonathan Hseu (@jhseu) profile picture

I manage the TensorFlow TPU team at Google Brain. I'm also working on a bot for Super Smash Brothers: Melee in my spare time: https://nitter.cf/t.co/S2eaRIZwDf

Joined June 2014
Jonathan Hseu retweeted
We just added a new tutorial that shows how to write Transformer (‘attention’ is all you need) in #TensorFlow 2.0 from scratch. Check it out here → goo.gl/6e9Xca
3
246
4
755
We're working on a new optimizer that seems to scale well: arxiv.org/abs/1904.00962
✨😍 Y'ALL: IT'S AN EAGER @TENSORFLOW BEAVER Artist credit: the endlessly talented @AshleyMcNamara. Isn't he the most adorable thing you've ever seen? His dam is even made of neural nets! (cc: @martin_wicke @edd @random_forests)
4
21
2
148
It is really great that TensorFlow is moving towards a public evolution/review process! e.g. github.com/tensorflow/commun…
1
35
1
174
Mesh TensorFlow has been open-sourced in Tensor2Tensor. It's a really promising direction for expressing model parallelism and training large models. github.com/tensorflow/tensor…
3
1
24
"Highly Scalable Deep Learning Training System with Mixed-Precision: Training ImageNet in Four Minutes," Jia and Song et al., Tencent: arxiv.org/abs/1807.11205
2
15
3
58
Jonathan Hseu retweeted
Our latest DAWNBench results are live: 8h52m for @TensorFlow to train ResNet-50 on ImageNet on a single @GCPcloud TPU (<$60), and just 30 minutes on half a TPU pod! dawn.cs.stanford.edu/benchma…
Jonathan Hseu retweeted
We just posted new DAWNBench results for ImageNet classification training time and cost using Google Cloud TPUs+AmoebaNet (architecture learned via evolutionary search). You can train a model to 93% top-5 accuracy in <7.5 hours for <$50. Results: dawn.cs.stanford.edu/benchma…
6
199
11
543