Placeholder Image

字幕列表 影片播放

  • [MUSIC PLAYING]

  • RICHARD WEI: Good morning, and welcome.

  • My name's Richard.

  • JAMES BRADBURY: I'm James.

  • RICHARD WEI: We're thrilled to talk to you

  • about a next-generation machine learning

  • platform, Swift for TensorFlow.

  • TensorFlow is a world-leading ecosystem for machine learning,

  • and we're always looking to make it better

  • for all kinds of users.

  • TensorFlow 2.0, available in alpha,

  • makes it easier than ever to get started with ML.

  • But we're also looking further beyond.

  • In this session, I'm going to tell you

  • about a project to rethink what a machine learning

  • framework could look like.

  • You might hear Swift for TensorFlow and think,

  • this is going to be about iOS apps.

  • But Swift is also a general purpose language

  • that we think will help enable the future of ML.

  • When figuring out what a next-generation machine

  • learning platform should look like,

  • we have been guided by three fundamental goals.

  • First, we would like to make sure

  • that every developer can be a machine learning developer.

  • Second, we'd like to make sure that every application can

  • benefit from the power of ML.

  • Third, we've seen that the frontiers of ML research are

  • pushing the boundaries of a software stack split between

  • Python and C++.

  • So we want a unified platform in the friendly and fast language.

  • This is why we decided to develop something

  • a little outside the box.

  • So we've taken the Swift programming language,

  • a general purpose and cross-platform language that

  • compiles to native code, and built general machine learning

  • functionality right into its core,

  • and combined its power with the TensorFlow ecosystem.

  • And this is Swift for TensorFlow.

  • Swift is a language designed to be easy to learn and use.

  • Let's take a look.

  • Swift's syntax is familiar and free of boilerplate.

  • Swift has types, so it catches errors early for you.

  • But it uses type inference, so it feels clean and simple.

  • Swift encourages fluent API design.

  • So your code is fluent and readable

  • at both the definitions side and the call side.

  • Swift is open source and has an open process

  • that allows us to propose new changes, new language features

  • for all Swift developers.

  • Swift supports prototyping platforms

  • such as a command line interpreter, Jupyter Notebooks,

  • and Playgrounds.

  • Just import your favorite libraries

  • and get started right away.

  • The TensorFlow library includes our carefully curated Tensor

  • APIs, layer APIs, as well as general abstractions

  • that help you build your own.

  • Now let's look at how to build a deep learning model in Swift.

  • This is an image classification model.

  • This is built out of several layers,

  • forming a simple convolutional neural network.

  • The call function applies every layer sequentially,

  • producing a final output.

  • Once we have defined the model, it's time to train it.

  • We start by initializing the model

  • in the stochastic gradient descent optimizer.

  • We'll be using random data here, x and y.

  • Now we write a training loop, apply the model,

  • compute the loss from labels using the softmaxCrossEntropy()

  • function, and compute the gradient of the loss with

  • respect to the model's parameters.

  • Finally, we can use the optimizer

  • to update all the differentiable variables and all parameters

  • in the model.

  • Our layer and optimizer APIs are inspired

  • by libraries like Keras, and are designed to be familiar.

  • A huge part of Swift for TensorFlow is workflow.

  • Lots and lots of research breakthroughs

  • in machine learning come from prototyping ideas.

  • Google Colab is a hosted Jupyter Notebook environment

  • that allows you to prototype your ideas right

  • in your browser.

  • A lot of researchers and developers

  • use Python with TensorFlow in Colab today.

  • And now we've made Colab work seamlessly with Swift.

  • Now let's show you Swift in action.

  • This is a Colab notebook.

  • As we start typing, we get an immediate code completion.

  • Here we use functional map to add 1

  • to each element in an array.

  • And it prints just like you expect.

  • Now let's look at a basic workflow in more detail.

  • Here's the dictionary of numbers.

  • And I want to find the largest even number in each array.

  • To do that, I'm using a for loop that loops through all elements

  • where the element is both greater than the largest--

  • than the last largest one found and is even.

  • Let's run that and see what happens.

  • Oops, what's wrong?

  • Well, it looks like Int doesn't have a property called isEven.

  • No problem, let's add one.

  • Even though the Int type is defined

  • in the standard library, we can still

  • add methods and properties to it.

  • In Swift, this is called an extension.

  • Here we define a Boolean computed property

  • called isEven, implement it in terms of the isMultiple method.

  • As you can see, Swift's syntax and naming conventions

  • look more like English than other programming languages.

  • And now everything works.

  • By importing TensorFlow, we get the power

  • of TensorFlow operators as functions and methods

  • on a Tensor type.

  • Many of these methods are the same

  • as they would be in NumPy and Python TensorFlow.

  • Other APIs make use of Swifty features,

  • like generics and argument labels.

  • TensorFlow also provides deep learning building blocks,

  • such as layers and optimizes.

  • Here's the same simple [INAUDIBLE]

  • that we just showed you.

  • We're on GPU, and it's running really fast.

  • Running the cells train the model.

  • It is that simple.

  • So that was Swift for TensofFlow in Colab.

  • Writing Swift for machine learning

  • is not all that different from writing dynamic languages.

  • With Colab, you can start Swift right away in your browser,

  • and it gives you code completion, free GPUs,

  • and you can even install a package, using Swift package

  • manager, right in your notebook.

  • You don't have to install anything.

  • It's really handy.

  • Next up, interoperability--

  • Swift interoperates transparently with C.

  • So you can literally just import a C header and use it.

  • But how about all the libraries--

  • everybody's favorite machine learning and data science

  • libraries in Python?

  • Do we have to read/write everything in Swift?

  • The answer is, we don't.

  • In Swift for TensorFlow, we have extended

  • Swift's seamless language interoperability to Python.

  • As you start using Swift for TensorFlow,

  • we want to make sure that you didn't miss all the utilities

  • and libraries that you're used to.

  • Now let's show you how interoperability

  • works in Swift.

  • Python interoperability is enabled by a lightweight Swift

  • library called Python.

  • Let's start by importing that.

  • Python.import lets you import any Python package installed

  • with pip.

  • Hundreds of popular packages are pre-installed in Colab.

  • Here we're going to use NumPy and Matplotlib.

  • We also enable their in-notebook visualization functionality.

  • Now we can call some common NumPy and plotting functions

  • directly from Swift exactly as if we're writing Python,

  • and it just works.

  • Now, OpenAI Gym is a Python library

  • of reinforcement learning environments, little games

  • for AIs.

  • Let's import that and solve one of the games.

  • One of these games is a cart-pole environment

  • where you train a neural net to balance a pendulum.

  • We'll use a simple two-layer neural network.

  • In each episode of the game is a loop.

  • We get observations from the Python environment

  • and call our Swift neural net to choose an action.

  • After running the game a few times,

  • we use the games that went well as training data

  • and update our neural net's parameters so that it will

  • pick better actions next time.

  • Repeating this process, we're getting better and better,

  • higher and higher rewards.

  • Finally, the problem's solved.

  • We can also plot the mean rewards.

  • As you see, you can mix and match

  • idiomatic Swift code, Python code, and a neural net defined

  • inline in Swift.

  • It is so seamless.

  • In fact, Python interoperability isn't hard-coded in the Swift

  • compiler.

  • It's actually written as a library in pure Swift.

  • And you can do the same--

  • do the same to make your libraries--

  • to make Swift interoperability work with Ruby or JavaScript

  • and other dynamic languages.

  • Now, in cutting-edge research and applications,

  • people often need to dive into low-level implementations

  • for better performance.

  • Swift's C interoperability makes that surprisingly easy.

  • Importing Glibc gives you access the C standard library from

  • qsort() all the way down to malloc() and free().

  • Here we allocate some memory using malloc(), store string,

  • and print it.

  • Now it's probably fair to say that Swift in Colab

  • gives you a better C prototyping experience than the C

  • language itself.

  • Now I hope you'll agree that the best parts of Swift

  • for TensorFlow is how easy it is to work

  • with code that is not in Swift.

  • And that is interoperability.

  • Having a few Python libraries in hand

  • is incredibly useful for day to day tasks in data science.

  • Swift for TensorFlow embraces interoperability

  • and lets you use C and Python without wrappers--

  • so seamless.

  • Now it's time to talk about the dark magic in Swift

  • for TensorFlow, differentiation.

  • It is fair to say that calculus is an integral part

  • of deep learning.

  • This is because machine learning models are really

  • just mathematical functions expressed in code.

  • But the special thing about them is that we don't just

  • want to evaluate those functions,

  • we also want to change the parameters,

  • or train them by changing their parameters based

  • on their derivatives.

  • A common technique adopted by deep learning frameworks

  • is called automatic differentiation.

  • Different from traditional programming,

  • doing calculus in code requires the language or the library

  • to understand the structure of your code.

  • But who to understand your code better than the compiler?

  • We think that computing derivatives from code

  • is not just a simple technique, but also a new programming

  • paradigm in the age of machine learning.

  • So we've integrated first-class automatic differentiation,

  • differentiable programming right into Swift.

  • First-class differential programming

  • means that native functions like this can be differentiated.

  • Here's a function of floats.

  • You can use a gradient operator to get its gradient function.

  • And just call it.

  • It's that simple.

  • And you can import TensorFlow, of course,

  • to compute gradients over tensors in the exact same way.

  • In fact, we've been using this feature in the model training

  • examples that you just saw.

  • Now instead of telling you how cool it is, let's show you.

  • Here's a simple function that works

  • with the native double type in Swift.

  • It squares x and adds y to it.

  • If we mark this function with differentiable,

  • the compiler checks to see if it can differentiate everything

  • that this function calls.

  • Now f is a differentiable function.

  • One of the things we can do with a differentiable function

  • is to get its gradient.

  • A gradient represents derivatives

  • of the output with respect to each of its inputs.

  • Here the derivatives are with--

  • here are the derivatives with respect to x and y,

  • produced by the gradient operator.

  • Well, sometimes it's more efficient to compute the output

  • and the gradient together.

  • Here's a function that calls the squaring function f.

  • We can use a value with gradient operator to do this.

  • All our differential operators like value and gradient

  • also work with arbitrary closures, not

  • just named functions.

  • Here we use Swift's trailing closure syntax to simplify it.

  • The closure represents the same g function passed directly

  • to the differential operator.

  • Now so far, we've seen differentiation on types

  • from the standard library and model training examples

  • earlier, using differentiation on tensors, layers, and models.

  • But in applications, we often need to define custom types,

  • like a point, which has an x and a y component.

  • Here we also define a custom function, dot(),

  • that takes the dot product of two points representing

  • vectors.

  • Now what if I call the differential operator

  • on the dot product function?

  • Boom, Swift reports an error at compile time saying point

  • is not differentiable, does not conform to the protocol

  • "differentiable."

  • Don't worry.

  • This is because Swift supports generics

  • and generic constraints.

  • The differential operator defined

  • for functions whose parameters and return type--

  • takes functions whose parameters' type and return

  • type conform to the differentiable protocol.

  • Now all we have to do is to make the point struct conform

  • to the differentiable protocol.

  • And now everything works.

  • Because a gradient operator returns a gradient tuple,

  • it expected two arguments, a and b.

  • We can simplify the code by using Swift's pattern matching

  • capability to retrieve the tuple elements in one line.

  • Swift's compile-time automatic differentiation

  • can also detect more subtle errors.

  • Here we define a function called roundedSum() on the point that

  • rounds the x and y components by converting them, first,

  • to integers, then converting the sum back to float,

  • back to a double.

  • Let's see if this function can be differentiated.

  • Well, it looks like we shouldn't have done that.

  • Functions over integers are not differentiable,

  • because the functions are not continuous.

  • And the compiler knows exactly why and tells you.

  • We showed you earlier that Swift has seamless interoperability

  • with C. But what happens if I use a differential

  • operator on a C function?

  • Well, let's try that.

  • Let's try to differentiate square root.

  • Well, it looks like it can't be differentiated.

  • This is because a function is not defined in Swift,

  • and the C compiler cannot compute derivatives.

  • But don't worry.

  • Let's write a Swift function that calls the C version

  • and write a custom derivative for it.

  • Now, in elementary calculus, the derivative of square root

  • is negative 1 over 2 root x.

  • We can define such a derivative as a Swift function that

  • returns both the original output and the closure that

  • applies the chain rule of differentiation

  • to compute the derivative.

  • If we mark this function as differentiating square root--

  • when we do this, we're telling the compiler, whenever

  • you differentiate a square root, treat the derivative function

  • as its derivative.

  • Now we can differentiate that function.

  • So this exact-- this is also exactly how

  • we made all the functions and methods in the TensorFlow

  • library differentiable.

  • It is that simple.

  • Isn't that amazing?

  • Having automatic differentiation built directly

  • into a general purpose language is unprecedented.

  • It's a super-exciting time to bring

  • your applications, intelligent applications,

  • to the next level.

  • Differentiable programming is new era of programming.

  • Language integration gives you useful compile-time errors

  • and the flexibility to make user-defined types

  • differential.

  • There's also a tremendous amount of depth here.

  • But if you're interested in learning how it works,

  • we have some detailed design documentation online.

  • Next up-- performance.

  • Swift is fast.

  • Swift has good low-level performance thanks

  • to its LLVM-powered compiler.

  • And it lets programmers use threads

  • for multi-core concurrency.

  • For deep learning code, we start with eager execution,

  • so writing TensorFlow code feels great.

  • But the compiler is also doing a lot

  • of things behind the scenes.

  • It's optimizing the tensor computation,

  • all without changing the programming model.

  • One trend we're seeing is that deep learning models

  • are getting integrated into bigger and bigger applications.

  • The problem is, doing this often requires

  • you to export models and treat them

  • as black boxes in your code.

  • But we think that Swift for TensorFlow can make it better.

  • Because your model is expressed just as code.

  • And it sits alongside the rest of the application.

  • And this enables all sorts of whole-program optimizations

  • that a compiler is really good at.

  • It makes debugging a lot easier.

  • What about performance in a research context?

  • Some of the most cutting-edge projects

  • like DeepMind's AlphaGo and AlphaZero

  • worked by bringing these three things together.

  • They use deep neural nets integrated into classical AI

  • search algorithms.

  • While it's possible for an advanced team like DeepMind

  • to build all of this, we think that an AI-first language

  • can make it easier.

  • And lowering the barriers in the infrastructure

  • between deep learning and traditional programming

  • can lead to new breakthroughs in science.

  • The MiniGo project is a ground-up, open source

  • replication of the central advances

  • of AlphaGo and AlphaZero using GCP and Cloud TPUs.

  • In order to achieve the best performance from the Monte

  • Carlo tree search algorithm that evaluates possible moves,

  • the author also had to rewrite everything in a complex mixture

  • of C++ and Python.

  • But Swift for TensorFlow gives you a better option.

  • Because MiniGo can be written in one language

  • without sacrificing performance.

  • Let's see what that looks like in Swift.

  • The Swift models repository on GitHub

  • contains machine learning models that we

  • built using Swift for TensorFlow,

  • such as common [INAUDIBLE],, transformer, and of course,

  • MiniGo.

  • On Linux or Mac OS, we can build a models repository

  • using Swift package manager via Swift build.

  • Or we can just open the IDE to develop machine learning models

  • like traditional applications.

  • The MiniGo game has three components--

  • one, the game board, two, the Monte Carlo tree search

  • algorithm, three, the convolutional model

  • to use with tree search.

  • They're all defined in Swift code in this workspace.

  • Recently, we've started doing some refactoring on the Monaco.

  • Now let's run some unit tests and see

  • if we have broken anything.

  • Whoa, what's wrong?

  • Well, it looks like the--

  • if we go to the test, looks like the shape of the value tensor

  • is expected to be 2, 1.

  • Well, where did that come from?

  • Let's just jump to the definition, the prediction

  • method that returned it.

  • Well, the prediction method simply

  • call self via the call method.

  • So let's look at that.

  • Now let's set a breakpoint and run the test again

  • to see what's going on.

  • At the breakpoint, we can print out

  • the shape of the value tensor using the lldb debugger,

  • exactly like application development.

  • Well, it looks like the tensor has shape 2, 1,

  • but the test expected just 2.

  • To fix that, we can simply flatten the tensor

  • by calling the flatten method.

  • Now let's drop the breakpoint and run everything again.

  • Great, all tests are passing.

  • We could run the game locally, but it'll

  • be pretty slow on a laptop CPU.

  • How about running it on a free GPU in a Colab.

  • Let's do it.

  • Let's open Chrome.

  • And here's a Colab that's installing the MiniGo

  • library that we've been working on using Swift package manager.

  • And so Swift package manager pulls directly from GitHub

  • and sets things up.

  • Now it's linking and initializing Swift.

  • Here we download some pre-trained weights using

  • some command line command.

  • And we also set up the game environment,

  • like the Go board and the players.

  • Now the game is starting to play.

  • The stones are being placed on the Go board one

  • step at a time.

  • Now we're running a Go AI on GPU in the Colab.

  • Isn't that amazing?

  • To advance through the game, here's

  • the same game that has been running for quite a while.

  • Isn't that cool?

  • So that is MiniGo and Colab.

  • Having deep learning models sit alongside complex algorithms

  • in an application and debugging through all of them--

  • the dream has finally come true.

  • Now you've seen how a unified programming model makes it

  • so easy to push the state-of-the-art in machine

  • learning research.

  • But we want to go even further.

  • For example, Google is making a major investment

  • in fundamental compiler technologies, an infrastructure

  • that will allow you to write custom kernels for hardware

  • acceleration right in Swift, and even in Colab someday.

  • Now James will talk to you about that work

  • and where Swift fits into the future of the TensorFlow

  • ecosystem.

  • James.

  • JAMES BRADBURY: Thanks, Richard.

  • So this is a broad-brush look at the lifecycle

  • of a TensorFlow model today.

  • You build a model, typically using Python.

  • And then you can run it in many different ways

  • on many different kinds of hardware.

  • And Swift can plug right into this ecosystem.

  • In fact, that's what we're doing.

  • We want it to be possible to use a TensorFlow model written

  • in Swift in all the places you can use a TensorFlow

  • model built in Python.

  • But the reality is there's a lot of complexity

  • here, many different hardware back ends and compiler stacks.

  • And thankfully, the team is working on a long-term project

  • to unify and simplify the situation with a new compiler

  • infrastructure called MLIR, or Multi-Level Intermediate

  • Representation.

  • Things like the TFLite model converter

  • and compilers for specialized hardware, all of them

  • will be able to share functionality,

  • like source location tracking.

  • And this will make life easier for everyone

  • in the ecosystem, all the way from app developers

  • who want a better experience like high-quality error

  • messages when exporting to mobile, all the way to hardware

  • vendors who want to implement TensorFlow support more easily.

  • Check out the TensorFlow blog for more details on MLIR.

  • Now the reason I'm bringing this up

  • is that MLIR is a great opportunity

  • to take advantage of the strengths of Swift

  • as a compiled language.

  • So while Python will be able to drive MLIR-based compilers

  • at runtime, Swift will also drive them

  • from its own compiler, bringing benefits

  • like type safety and static binaries.

  • In short, for users who need the flexibility,

  • we can make the entire TensorFlow ecosystem directly

  • accessible from Swift, all the way down to low-level device

  • capabilities for writing custom kernels, all

  • the while improving the experience for everyone who

  • uses TensorFlow.

  • So that's MLIR.

  • Let's talk about the original vision again.

  • We think Swift for TensorFlow can help every developer become

  • a machine learning developer.

  • Thanks to Swift's support for quality tooling like context

  • aware, auto-complete, and the integration

  • of differentiable programming deep into the language,

  • Swift for TensorFlow can be one of the most productive ways

  • to get started learning the fundamentals of machine

  • learning.

  • In fact, you don't have to take our word for it.

  • Jeremy Howard's fast.ai recently announced

  • that they're including Swift for TensorFlow

  • in the next iteration of their popular deep learning course.

  • We think Swift for TensorFlow can

  • help make sure every app can benefit

  • from the power of machine learning

  • by bridging the gap between application code and machine

  • learning code.

  • And we think that Swift for TensorFlow

  • will provide a great option for researchers

  • who need a high-performance host language like Swift.

  • To break down the barriers for you to get started,

  • Google Colab hosted notebooks give you

  • free access to GPU resources right in your browser.

  • And Jupyter Notebooks let you quickly prototype your ideas

  • locally or on remote machines.

  • Swift for TensorFlow also supports Linux and Mac OS

  • natively.

  • So you can use your favorite IDE or run

  • the interpreter in a terminal.

  • RICHARD WEI: Swift for TensorFlow's 0.3 release

  • is available today, which contains

  • all the technologies that powered the demos

  • that you just saw.

  • We think it's ready for pioneers,

  • especially if you're running into limits of Python

  • as a host language or mixing deep learning

  • with traditional programming.

  • JAMES BRADBURY: Swift for TensorFlow

  • is not yet production-ready.

  • But if you're excited about what you see,

  • we have Colab tutorials you can run in your browser and binary

  • tool chains you can download and use.

  • We're actively designing and building high-level machine

  • learning libraries and improving the programming experience.

  • RICHARD WEI: The whole project is developed in open source

  • at github.com/tensorflow/swift.

  • Try it out today, and we think you're going to love it.

  • JAMES BRADBURY: This is an incredible opportunity

  • for developers like you to help shape the future of TensorFlow.

  • RICHARD WEI: Thank you for coming, and have a great I/O.

  • [MUSIC PLAYING]

[MUSIC PLAYING]

字幕與單字

單字即點即查 點擊單字可以查詢單字解釋

B1 中級

Swift for TensorFlow (Google I/O'19) (Swift for TensorFlow (Google I/O'19))

  • 3 0
    林宜悉 發佈於 2021 年 01 月 14 日
影片單字