this time to talk about the TensorFlow eager execution

runtime.

This is a very broad topic and there

are lots and lots of things we could cover,

so I'm going to lightly graze many, many different parts

of our code base.

I'll give you a lot of function names

and file names and things like that

that you can use to familiarize yourself with something,

but if you're in the room right now, by all means,

ask questions.

Like, this is-- there's some buffer time at the end

to account for variability here.

And I think the more we can maximize shared understanding

of this stuff, the better.

So the way I thought we could go about this

is to do a very, very deep dive on what

actually happens in TensorFlow, starting from TF2--

when you type a very simple line of code, in this case

a tf.nn.relu off some Python numbers.

And I think if you were to start doing this, probably

the first thing you'd do is graph the TensorFlow code

base to find where would we define ReLU.

And if you do that, you will find

that we have some definitions of ReLU in Keras,

but you won't find a single definition of ReLU itself

in core TensorFlow, and that might be

a little surprising at first.

It might put a damper on this whole,

let's find out what actually happens when we run ReLU

business, but the way ReLU comes from is because it's enough

that we've implemented in C++ and we didn't need to put

a complicated Python API around it.

We can just generate the Python code to call ReLU.

So the way it's actually defined,

it's defined using the same mechanism

we use to register all the ops in TensorFlow.

So for every core operation in TensorFlow

that's visible to the runtime, we

have a registration that looks like this--

REGISTER_OP.

It takes a name and you say how many inputs, how many outputs,

what attributes it has.

If you want to know more about attributes-- what things

are allowed in there-- they are how

we can make our ops polymorphic and have the same operation

have different types of outputs, so different numbers of outputs

and things like that.

There is a lot of documentation about this in TensorFlow.org,

if you search for how to define a new op.

And another interesting thing in there

is that we also register a shape_inference function.

ReLU, thankfully, is one of the simplest

ops we have-- it just has one input, one output,

they have the same shape.

So we can use a pre-built shape_inference function

that just says the shape does not change.

Other ops will have vastly more complicated

shape_inference functions.

And the nice thing is that we can

run these functions offline for our building graphs

without actually having the values of any tensors

and still be able to prove things

about the shapes of intermediate tenses and outputs

of your computation.

This is maybe the best tool we have for catching bugs now.

So if you want to look at the shape_inference code,

that's where you'd hook into.

Now that we have that registration,

we run some complicated code that generates Python code

to actually call ReLU.

And if you look in bazel-genfiles, files,

you will find a file named gen_nn_ops.py and this file has

the actual def ReLU that we call.

And as you can see, it's not pretty

and there's a lot of stuff going in there.

The first line deals are dispatching

so that we can define ReLU not just for normal tensors,

but also optionally for sparse tensors

and ragged tensors and other composite types.

The second line has tf_export and what this does

is define the TensorFlow public API.

Every symbol that you get when you are using TensorFlow by tf

dot something is defined somewhere

from a tf_export decorator like this one.

There will be a future video on how exactly this works

and why we do things this way instead

of relying on Python's normal, you know,

name spacing mechanism.

But you can probably guess that it's because TensorFlow

is very complicated.

But essentially, you'll see this,

and this generated code for ReLU has a bunch of cases in it.

That are roughly four.

You have an eager fast path, an eager slow path,

you have a graph mode path, and kind of a side hook

for the symbolic execution.

But here, let's focus on the eager paths.

In the first one, the first thing

that we're actually doing here is

we're checking to see if we're in eager mode or not.

And to do that, we look at this context thing.

This context thing is part of the core of the TensorFlow v 2

runtime.

It's the moral equivalent to the session,

but it's longer lived than the session

and represents more things.

So what is it?

From Python, the context is this class

that's defined in a file called [? context.ui. ?]

And it's a collection of a lot of things

that your Python program needs to be aware to connect

to the TensorFlow runtime.

It stores things like, am I in eager mode or in graph mode?

Or if someone used a with tf.device decorator, what

device am I supposed to be executing code in?

And it stores things like, what's

your name scope and many other things.

And all of this information that the context

stores, in general--

the things that can change during the execution

of a program--

they're all stored in ThreadLocal stacks.

Usually stacks, because we have these nested things

like with tf.device, with tf.device, with tf.device,

so you'd like to be able to pop the stack

to go back to where you were.

And ThreadLocal because it's very important to us

that a TensorFlow runtime itself be thread agnostic,

so that if you write two threads and one is doing

a reinforcement learning learner and the other's doing

an agent that's talking to some game, when the agent wants

to use its GPU, it shouldn't necessarily make the learner

use the U, and vice versa.

Providing some kind of resolution between the threads

is what we felt like was the right way,

so that at least each single thread

can feel like it's its own single-threaded Python program.

We use this a lot in distribution strategies,

like MirroredStrategy uses a lot of threads under the hood,

so it's really important that things are thread-safe.

In the Python context, essentially it mostly is

a wrapper around the C++ context.

It's available for the TensorFlow C API.

And this is the core thing.

It has a bunch more methods than just these--

like, you can do a lot more things

than just listing devices.

One thing I'd like to call out is that right now,

there are some things that are done in Python, like storing,

whether in eager mode and graph mode,

and some things that are done in C++.

And it's not the set of things that are done Python and set

of things that are done in C++ are likely to change.

I think as TensorFlow evolves, more and more things should

migrate from the Python context to the C++ context,

which will make things more language agnostic,

but also faster.

And you know, if everything in the context was in C++,

then all the generated Python code could have just been C++

code and it would have to--

we'd be able to get out of the overhead of executing Python

much sooner and remove performance problems

in our APIs.

So once you know you're in eager mode,

we try to do this fast path execution.

In this, the fast path is some complicated C code

that mostly does the same things that the fallback case is

trying to do.

So I don't think it's necessarily worth reading that.

I would rather look at the simpler

code and the fallback path.

And it's here.

So this is where we actually implement the ReLU function.

And what are the interesting things here?

First, we have this function called args_to_matching_eager,

and what it does is it takes a bunch of tensors,

a bunch of things that can be tensors,

converts them all to tensors of the same dtype.

And in the case of ReLU, there is only one.

But in the case of any other ops, like Add or MatMul,

they take multiple inputs, some of which might be tensors,

others might be numpy.arrays or Python lists or variables

or other objects that you can convert to a tensor,

but they're not necessarily tensors.

And this is the thing that's responsible for canonicalizing

everything.

Then once we have--

but here you might want to stop me a little and ask,

what is a tensor at the level of a TensorFlow runtime?

And the eager tensor class is a thing

that's half implemented in Python

and half implemented in the Python C API,

but it's a relatively thin wrapper around this thing

that we call TensorHandle that you can see in our TensorFlow C

API.

And this TensorHandle, it represents a potentially yet

to be computer tensor which is going to live on some device.

And we know what device it's going

to live on, because we know what device we executed

the operation that generates the tensor on.

And there are roughly three things

you can do to a TensorHandle, really.

You can ask about its shape and dtype and stuff-- that's one.

You can give it to execute--

that's another one.

And you can copy it to a device.

Also, I guess there are four things,

and the fourth thing is you can ask, hey, what is its value?

And when you want to force the evaluation of its value,

the TensorFlow runtime might have

to pause until it can give you the result.

You might need to copy the tensor from some remote device

and do all sorts of other things, but once you're done,

you get a TF_Tensor, which is essentially

just a pointer to some data.

But every other operation that is not

looking at the value of the tensor

is something that the runtime is free to delay, reorder,

as long as it respects the intended

semantics of your program.

And this is very important, because when

you're working with things like GPUs and TPUs,

you'd like to be able to dispatch operations that

are running the hardware as quickly as you can

in a way that's asynchronous.

If the Python thread can race ahead of the GPU

as it's executing operations, we can fully

utilize the GPU hardware and get the maximum performance.

And even if the eager runtime, we

can do this if the GPU kernels are sufficiently heavy.

There's a few cases in which we need to copy tensors,

even if they're local on the CPU.

And some of those cases are going away, like currently,

we copy string tensors every time we try to look

at their values because TensorFlow's internal string

representation is a complicated C++ class and we'd like to have