I think by end of next class I would have fallen back in love with programming after a gap of 10 years. What a smasher this lecture has been! Thank you @jeremy and @clattner and team!
14 Likes
If I use S4TF to build and train a model, is there a way to deploy the model to run on iOS? What would be needed for this?
2 Likes
Do you have any plan to support functions for data I/O especially for image loading and dumping?
Deep Learning, Swift, Python… The greatest things I’ve been working with recent years now going to come together into a single ecosystem. That’s amazing. The world of Machine Learning and numerical programming changes in right direction, I believe.
2 Likes
To people being excited about learning Swift and S4TF and even awesome low level stuff that were discussed in this lesson, please please please consider sharing what you’ve learned and how you’ve learned it on blog posts and share it here on the forums!
10 Likes
There are multiple possible ways, but one way we would really love to see happen is integrating the model as native Swift code in any mobile app. This will require a good compiler infrastructure and a good runtime that can optimize models’ code size and performance.
In the near term, it will be possible to export models as TensorFlow graphs, and convert them to TensorFlow Lite or Core ML models using their conversion tools. Relatedly, parts of TensorFlow Lite will be powered by MLIR very soon!
5 Likes
I am trying to find some time to just learn some of these things before I can even start writing some posts 
Lot’s of information these weeks!
6 Likes
S4TF as an AI-first language? 
2 Likes
You can check out our tutorials!
If you have time for just one, check out our end-to-end model training tutorial.
I’d also highly recommend our custom differentiation tutorial. Differentiation in Swift is first-class and customizable, but APIs are different than other systems. (More info on differentiation coming up in the 4/30 lesson!)
10 Likes
Thanks to the S4TF team for dropping by to answer questions! There’s a ton of material here to keep me busy for a good while!
10 Likes
Are there any projects currently going on to be able to inject Swift into Python code? It’s really cool that you can inject Python into Swift, but if you could do the opposite, that would be a great way to get the speeds that Swift provides and still stay in Python.
1 Like
I note that the
.../anaconda3/envs/s4tf-a/lib/libuuid.so.1 no version information availablecompiler message is not generated by your notebooks when activating the%install cell. I have this issue.
I have gone through the prerequisites for the ubuntu18.04 platform
clang etcwhich I have installed on my desktop. I have been using your nightly script for updating s4tf.
I have googled this issue but can only find a generic explanation and that a possible resolution is to rebuild the library which is not a realistic one. Another was to rename the version part of the name which was poo poo’d in respect of the library rebuild
If you have any insights into this issue I would grateful .
So far to resolve this I have created a python environment s4tf-a without fastai in which libuuid 1.0.3 is installed but when I ls the environment location it is at libuuid.so.1.0.0. Also the notebooks work despite this behaviour
On slide #10, the acronym GIL is defined here: https://wiki.python.org/moin/GlobalInterpreterLock
On slide #50 (towards the end of this lesson), the Halide video has a great example of showing several ways you could compute the blur of a photo. Although the video is about a language called Halide, it is a good illustration the issues it raises are universal:
- there are different choices in how a given matrix operation is implemented
- these choices impact speed and memory usage
- these different approaches have trade-offs
Lots of interesting idea! Also excited about the direction we’re moving towards with stuffs like MLIR is the ability to have Domain Specific Languages (DSL) where you could write for example, Halide directly.
1 Like
This is a great point! So far, Python interoperability has focused on calling Python from Swift. But calling Swift from Python is interesting because it enables people to import Swift into existing Python codebases: to write Swift models and to use Swift features like differentiation.
The S4TF team is aware of the “calling Swift from Python” direction but no one is actively working on it - it’s a great topic for exploration!
Technically, it should be possible by using the @_cdecl attribute in Swift to declare C-compatible functions, then importing those functions into Python. Here’s a crazy project leveraging @_cdecl to write custom PostgreSQL types/functions in Swift, sharing to spark inspiration!
6 Likes
Yes, first-class language support for machine learning!
I believe that requires:
- Automatic differentiation, necessary for gradient-based optimization. Check out what’s possible in Swift today! More info coming up on 4/30.
- Compiling code for accelerators (e.g. GPUs and TPUs).
4 Likes
Watched probably first 10 minutes of the video and I’m already sold for Swift. Either I’m brainwashed or it is the future.
5 Likes
A team of really incredible people pulled together to make it happen!
-Chris
7 Likes
this is an amazing initiative and collaboration - thx