Swift API

[janerivi]

Swift is a very popular and expressive language with a large community of pro developers.

[ghost]

Providing APIs for every 'very popular' lang would probably bloat the main project quite quickly.

[janerivi]

I understand the sentiment, but Swift, C# and Java are not "every" language, and this is not "every project" either. This is open source and Github so there might be enough hands to man the maintanance of ports to the most popular languages for such an important piece of software. Python is quite popular in academia but a little less popular with the bulk of private sector developers.

[davidzchen]

Given that TensorFlow is designed to also run on mobile and that Apple is recommending Swift as a primary language for iOS development, I think having a Swift API for TensorFlow would be a good idea.

I don't think adding support for the most widely-used languages such as Java, C#, and Swift would be a bad idea. The gRPC project already supports 10 languages, including Java, C#, and JavaScript (via Node.js).

@jamesliu96 Feel free to open an issue for adding a Go API.

[vincentvanhoucke]

A Swift frontend, especially for simply running graphs for inference on mobile would be great. I don't know of current plans to provide one, so feel free to dive in. Since our exact plans on accepting external contributions are still in flux, it would be a good idea to check in with the discuss mailing list with a draft of the code ahead of time to figure out where it should live exactly.

[loopbum]

I couldn't agree more with janerivi and davidzchen. Swift has many advanced features that make it one, if not the best choice for full API support for TensorFlow.

This is especially true since Swift is now open source. This language is very pleasant to work with and has all of the great features of top expressive languages like Haskell; Generics, Closures, etc.

I would not agree that C# or Java are good candidates in terms of speed. Glyn Williams is a trusted and respected contributor and writer in many areas and had this to say at:
https://www.quora.com/In-terms-of-performance-speed-is-Swift-faster-than-Java

Glyn Williams:
Swift is significantly faster than Java.
It compiles to native code. And a number of the language features enable an optimizing compiler to produce very fast code indeed.

Vivian Keating:
Swift will typically run faster than Java, since it compiles into native code via the LLVM compiler. Most recent metrics find it comparable to C++ in production.

This is the tip of the iceberg in terms of comparisons. Apple took great care in designing this modern language and how it compiles to native code, and now it's Open Source.

[rxwei]

I'll take on this.

[johndpope]

I have this swift docker file with parsley mcparsefsce / tensorflow + separate swift 3 dev container. - I will plugin swift grpc once google resolve an issue with their just released grpc library

https://github.com/johndpope/DockerParseyMcParsefaceAPI

I could use some help to get it all working.

Some @nubbel stubbed out proto files in swift here
https://github.com/nubbel/swift-tensorflow/tree/master/types
Should be able to leverage this environment to get above working.

[HelloFillip]

I have been working on a painful private project for a while using Swift. I'll be migrating the project API library to Swift 3 soon and opening it up under the same licensing as TensorFlow (Apache 2.0). It was reasonably complete before Swift 3 and I plan to have the Swift 3 version complete before the end of 2016, sans mistakes from original, and including lessons learned.

[johndpope]

@siilime fyi - Google released an official swift3 grpc library last week.
https://github.com/grpc/grpc-swift
There's still some heavy lifting to do to get it all working(eg generate swift 3 services classes). You may want to revisit your implementation and yield to official library. Surely less painful in the long run. I'm aware of some other grpc libraries and wrapping the c classes into swift module.(just search on GitHub (swift + grpc) there's one by Huawei. My interest is to get it working server side. If anyone wants to help / check the issues grpc-swift for docker issues ticket.

[HelloFillip]

@johndpope Good tip. Thanks. My focus was on the server side originally, and it's where the majority of my work has been so I'll be taking a look at that.

[johndpope]

Looks like @rxwei has made some progress here around providing some basic test cases around a swift wrapper for the tensorflow c api.

https://github.com/rxwei/tensorflow/tree/master/tensorflow/swift
https://github.com/rxwei/CTensorFlow

https://github.com/tensorflow/tensorflow/blob/master/tensorflow/c/c_api.h

[HelloFillip]

I've done a lot of review work over my original code base and the official Swift gRPC code base in the past few days, and although there's advantages to using the gRPC in some cases, there's still a lot missing, (specifically an easy option to train TensorFlow directly from Swift), and those gaps are the areas I'll be focusing on, whilst aiming to develop a complete solution in Swift.

Good start with @rxwei contribution so will aim to assist in that, whilst migrating existing code base.

[rxwei]

Currently the C API is under-documented and offers only a few core components, Graph, Tensor, and Session. I start to wonder if wrapping all the C API is worthwhile.

[HelloFillip]

@rxwei What feasible alternative approaches are there?

[johndpope]

for reference - found this - https://github.com/somaticio/tensorflow.rb
a ruby wrapper around tensorflow. (there maybe more)
It looks like they wrapped the c api in another c wrapper. Not sure if this helps.
https://github.com/somaticio/tensorflow.rb/tree/master/ext/sciruby/tensorflow_c/files

Another shot is to use grpc to interact with tensor.
but the services layer needs fleshing out which will be simpler once google can spit out service class definitions (think afnetworking for protobuf) for grpc

the python API has hundreds of granular parameters.
https://www.tensorflow.org/versions/r0.10/api_docs/python/index.html

compared to c++ API / c api.
https://www.tensorflow.org/versions/r0.10/api_docs/cc/index.html
Env

tensorflow::Env
tensorflow::RandomAccessFile
tensorflow::WritableFile
tensorflow::EnvWrapper
Session

tensorflow::Session
tensorflow::SessionOptions
Status

tensorflow::Status
tensorflow::Status::State
Tensor

tensorflow::Tensor
tensorflow::TensorShape
tensorflow::TensorShapeDim
tensorflow::TensorShapeUtils
tensorflow::PartialTensorShape
tensorflow::PartialTensorShapeUtils
Thread

tensorflow::Thread
tensorflow::ThreadOptions

If you're looking to machine learning in swift today (nov-2016) it'd pay to shop around for other libraries. https://github.com/search?utf8=%E2%9C%93&q=swift+ml&type=Repositories&ref=searchresults

One day - another (somewhat unfeasible) approach could be to rewrite the python library in swift.

[rxwei]

@johndpope Honestly none of those Swift ML libraries are nearly ready, cross-platform, nor written in Swift 3.0+.

[rxwei]

@johndpope service approach makes sense for now, as a wrapper.

[johndpope]

related grpc/grpc-swift#2
also there is this swift library for those that want to get their feet? / toes wet. https://github.com/qoncept/TensorSwift

[rxwei]

@johndpope That one is dependent on Darwin platforms and Apple's Accelerate framework. Not suitable for general use.