Add tensorflow

[Bidanets]

Please, add tensorflow.

[holzschu]

Noted, and added to the list of features: https://github.com/holzschu/Carnets/projects/1 It seems to require keras, which is already on the list; and it seems to require its own compiler, which could make cross-compiling a bit difficult. But it's now on the list.

[Bidanets]

On the contrary, keras package requires tensorflow background. (tensorflow or theano background) -> keras So, tensorflow is independent of keras.

[holzschu]

Thanks for the clarification. I made a note of that. It is still difficult to cross-compile, and even more so with custom compilers, so it’s not going to be instantaneous.

[Bidanets]

In this article https://preslav.me/2018/08/12/data-science-on-your-ipad/ about Pythonista 3 there are words: "Pythonista comes pre-packed with a ton of libraries we love and use every day: NumPy, Matplotlib, Requests, BeatufulSoup, etc. Yet, lacks some of our favourites, due to Apple’s restrictions: Pandas, SciPy, Scikit-Learn, and unfortunately, TensorFlow."

[Rui-K]

Really need tensorflow too, keep working on it please!

[kkebo]

Recently, Apple has released Mac-optimized TensorFlow. It uses ML Compute framework as a backend of TensorFlow and the framework also supports iOS 14+. I think this variant of TensorFlow is more suitable for Carnets than the normal one.

[CaptainDario]

@holzschu Are there any news regarding this feature? Maybe as @kkk669 stated it will be easier with the MacOS tensorflow port?

I think having tensorflow/keras for iOS would make the iPad much more "pro".

[holzschu]

The current TestFlight version has CoreMLTools, which is not tensorflow, but a step in the right direction. I'm working on the rest.

[essandess]

Thanks for the clarification. I made a note of that. It is still difficult to cross-compile, and even more so with custom compilers, so it’s not going to be instantaneous.

I’d also love to see TF.

I recommend using the Apple-compiled tensorflow-macos and tensorflow-metal packages, which should run quite fast on the latest iPad Pro’s with M1 and M2 chips. These can be downloaded directly from PyPi.

Rats—based on this comment, this approach won’t work:

Dynamic libraries contain specific code to prevent use on a different platform (you can't use Arm64 OSX libraries on Arm64 iOS phones). https://github.com/holzschu/Carnets/issues/245#issuecomment-1083314362

I’d also like to see PyTorch with MPS support. An example build formula for macOS is at MacPorts.

[Marcelo-GG]

Another vote for TensorFlow

[Air-Craft]

still no tensorflow :/. +1 anyway :)

[Knight7561]

+1