Swift for TensorFlow: The Next-Generation Machine Learning Framework (TF Dev Summit '19)

[@aadimator]

I'm also super excited for Swift + Fastai course. Looking forward to your lectures there Chris.

[@sucim]

Chris Lattner is this MVP kind of guy

[@tissuebox1229]

holy crap this is amazing

[@martinlasek]

I haven't worked with neural networks but this talk was still super interesting stuff! The python interoperability 😳! The idea of the custom debugger is awesome, too! I liked it 😊!

[@aadimator]

Guess I have to learn Swift now. It would've been really exciting if Swift was supported officially on Windows as well. Now, I'll have to set up a Linux Machine for this. Still, excited about learning new things.

[@OttoFazzl]

Even though it all sounds really interesting, it still at a very early stage. I suggest you follow the fastai part 2 course that will come out in June that will cover Swift and explain in much more detail how it will be useful for deep learning.

[@RuslanLagashkin]

Syntax looks to be rust-inspired. Loving it.

[@karanbirchahal3268]

Would love for Rust to do something like this !

[@ОлегНерж]

wow ! it is... woooow

[@matthewygf]

anyone understand the part where he is simulating data parallel ? from what i understand , data parallel doing grad at the same time at other devices, i mean shouldn't that be using some sort of multi-threading and average it .... or asynchronous update ?

[@MarkBrowning]

The tiered batches aren't simulating the *performance* of a cluster of GPUs, they were simulating the results: part of the issue of large batch sizes is the reduction/normalization across the whole batch. Here, they showed just addition of batch gradients without normalization, and implied you could play around with other normalization techniques or model updates (skip connections, etc)

[@vaibhavbv3409]

is gpu support still there

[@dimka11ggg]

No Nvidia web driver, no gpu

[@softwareinternals]

at 22:46 the x component of the gradient for cell output 19 should be ~17.12, you can verify at colab.research.google.com/github/saeta/s4tf-dev-summit-19/blob/master/TF_Dev_Summit_2019_S4TF_Prezo_03_Differential_Programming.ipynb#scrollTo=v9vRJK-mhFFp

[@lionardo]

but why swift? swift is not even platform agnostic as of now. If speed and performance is necessary you can use Julia.

[@tbass134]

RIP Swift for Tensorflow

[@proweiqi]

Double bias is such a dumb idea. I know it's for a demo, but it's the same as one bias but higher variance.

[@mahmud-i9c]

i dont like apple, but i love swift

[@EzekielPrellus]

Biggest engineering problem here is that "Swift for TensorFlow" is an unwieldy mouthful to pronounce. Needs a map-reduction to STF.

[@AdrianStabiszewski]

I hope they implement this also with Dart. Together with Flutter would be a perfect fit for mobile development.

[@deep.space.12]

Dev A: How do we easily get people started on TensorFlow? Dev B: Let's add another language! It will surely make things easier! Dev A: Just like tfjs we're supporting a popular use case? Dev B: Nah. We are making them learn a brand new language because it has _types._

[@samarioantonio]

import Python lol nice

[@proweiqi]

The language looks to be TOO verbose! I think Julia is the way to go.

[@buivietdung1995]

Yeah, I was surprised by the decision, at first I thought Julia would be the winner, hand down. Well, anyhow, Swift is not the bad choice, let see how it goes

[@navneetkrc]

The most important thing is community, FASTAI fellows will be contributing a lot for swift4TF. Watch this go big. Chris Lattner himself will be teaching in 2 lectures this year in fast.ai MOOC course, that is going to be released this June.