Get a taste of protocol-oriented differentiable programming.

This repository hosts Swift for TensorFlow's deep learning library, available both as a part of Swift for TensorFlow toolchains and as a Swift package.

This library is being automatically integrated in Swift for TensorFlow toolchains. You do not need to add this library as a Swift Package Manager dependency.

Open an empty Colaboratory now to try out Swift, TensorFlow, differentiable programming, and deep learning.

For detailed usage and troubleshooting, see Usage on the Swift for TensorFlow project homepage.

Simply import TensorFlow to get the full power of TensorFlow.

import TensorFlow let hiddenSize: Int = 10 struct Model: Layer { var layer1 = Dense<Float>(inputSize: 4, outputSize: hiddenSize, activation: relu) var layer2 = Dense<Float>(inputSize: hiddenSize, outputSize: hiddenSize, activation: relu) var layer3 = Dense<Float>(inputSize: hiddenSize, outputSize: 3, activation: identity) @differentiable func callAsFunction(_ input: Tensor<Float>) -> Tensor<Float> { return input.sequenced(through: layer1, layer2, layer3) } }

var classifier = Model() let optimizer = SGD(for: classifier, learningRate: 0.02) Context.local.learningPhase = .training // Dummy data. let x: Tensor<Float> = Tensor(randomNormal: [100, 4]) let y: Tensor<Int32> = Tensor(randomUniform: [100])

One way to define a training epoch is to use the gradient(at:in:) function.

for _ in 0..<1000 { let 𝛁model = gradient(at: classifier) { classifier -> Tensor<Float> in let ŷ = classifier(x) let loss = softmaxCrossEntropy(logits: ŷ, labels: y) print("Loss: \(loss)") return loss } optimizer.update(&classifier, along: 𝛁model) }

Another way is to make use of methods on Differentiable or Layer that produce a backpropagation function. This allows you to compose your derivative computation with great flexibility.

for _ in 0..<1000 { let (ŷ, backprop) = classifier.appliedForBackpropagation(to: x) let (loss, 𝛁ŷ) = valueWithGradient(at: ŷ) { ŷ in softmaxCrossEntropy(logits: ŷ, labels: y) } print("Model output: \(ŷ), Loss: \(loss)") let (𝛁model, _) = backprop(𝛁ŷ) optimizer.update(&classifier, along: 𝛁model) }

For more models, go to tensorflow/swift-models.

• Swift for TensorFlow toolchain.
• An environment that can run the Swift for TensorFlow toolchains: Ubuntu 18.04, macOS with Xcode 10, or Windows 10.
• Bazel. This can be installed manually or with Bazelisk. You will need a version supported by TensorFlow (between _TF_MIN_BAZEL_VERSION and _TF_MAX_BAZEL_VERSION as specified in tensorflow/configure.py).
• Python3 with numpy.
• CMake. CMake 3.16 or newer is required to build with CMake.

Note: Building with SwiftPM does not include changes to X10 modules.

$ swift build

Note: Testing with SwiftPM does not run X10 tests.

$ swift test

Note: CMake is required for building X10 modules.

In-tree builds are not supported.

Note: To enable CUDA support, run export TF_NEED_CUDA=1 before the steps below.

Note: If swiftc is not in your PATH, you must specify the path to it using -D CMAKE_Swift_COMPILER=.

This will build X10 as part of the build. Ensure that you do not have the x10 modules in the toolchain that you are using to develop here.

cmake -B out -G Ninja -S swift-apis cmake --build out

To run tests:

Note: To view failure output, run export CTEST_OUTPUT_ON_FAILURE=1 before running tests.

cmake --build out --target test

If you are not intending to develop X10, you can reduce the build times by using the bundled X10 in the toolchain using -D USE_BUNDLED_X10=YES -D USE_BUNDLED_CTENSORFLOW=YES:

cmake -B out -D USE_BUNDLED_CTENSORFLOW=YES -D USE_BUNDLED_X10=YES -G Ninja -S swift-apis cmake --build out cmake --build out --target test

On macOS, passing -D BUILD_TESTING=NO is currently necessary to skip building tests. This avoids an error: cannot load underlying module for 'XCTest'.

cmake -B out -D USE_BUNDLED_CTENSORFLOW=YES -D USE_BUNDLED_X10=YES -D BUILD_TESTING=NO -G Ninja -S swift-apis cmake --build out

Please report bugs and feature requests using GitHub issues in this repository.

Discussion about Swift for TensorFlow happens on the swift@tensorflow.org mailing list.

We welcome contributions: please read the Contributor Guide to get started. It's always a good idea to discuss your plans on the mailing list before making any major submissions.

In the interest of fostering an open and welcoming environment, we as contributors and maintainers pledge to making participation in our project and our community a harassment-free experience for everyone, regardless of age, body size, disability, ethnicity, gender identity and expression, level of experience, education, socio-economic status, nationality, personal appearance, race, religion, or sexual identity and orientation.

The Swift for TensorFlow community is guided by our Code of Conduct, which we encourage everybody to read before participating.