PAIGE BAILEY: Hi, my name is Paige Bailey.

And I am here today to deliver the talk of Karmel

Allison and Yash Katariya, who unfortunately couldn't be here

today.

But for the purposes of this talk,

just imagine that I am Karmel.

So to get started--

my son is in kindergarten.

And he's just starting to learn to read.

And one of my favorite parts of this process

is that even before he could reliably decode words,

he started to produce them.

And, of course, doing what I do, I

can't help but be amazed at how similar his learning

process is to neural nets.

The rule set for spelling and grammar

is so complex in English that we rarely present

a set of instructions that he can follow.

Instead, he starts almost randomly with letters

and sounds.

And then he gets feedback from me

and from other readers in his life which he incorporates.

Some things he memorizes.

Some are lucky guesses.

And somehow over months of learning,

he started to form a consistently interpretable

mental model of written words.

And he produces mostly understandable text.

So here you can see, "I was eating breakfast

with my cousins and my sister."

I've been particularly fascinated by my son's learning

here, because it happens to align with what many of us

have started to call an NLP revolution.

We have enough data and enough tools

that we've started to rapidly push the cutting

edge in natural language understanding and related

tasks.

I've heard it said that, with text today,

we're at a cambrian explosion, like when

ResNet was first published.

We're at the beginning of this faster than--

faster than expected progression of language models.

So we're going to take advantage of all of this research

and tooling.

And we're going to teach a simple neural net to generate

the next word of a phrase.

And we'll use this to tell a robot children's story.

So the first thing we need is data.

And we're going to take advantage

of the children's book test corpus released by Facebook.

This data set is a set of Creative Commons

children's books that have been converted into a series

of passages and fill-in-the-blank style

questions about each passage.

From this model, we just need the raw book text which

looks like what you see here.

As an aside, these books are out of copyright,

which means they're often old.

And that means the corpus is full of literature

that's problematic by today's standards.

I'm going to gloss over that for the sake of this talk.

But if you go and actually use this data set,

please do consider what cultural norms

you might be teaching your machine learning models.

So first, we're going to load the data.

Now, that we have a corpus, we can

load into our Python interpreter and start playing with it.

Here, we use the text line data set to load the data.

And we can simply print out a few lines to see what we have.

So tell me what you notice here.

It looks like we have some cleaning work to do.

So using our data set much like we would use a NumPy array

or Pandas data frame, we can filter

and map transformations across it.

Here, I'm dropping those pesky book titles

and filtering out punctuation within each row.

And after that transforming, I can print out a few lines

to check and make sure the data looks as expected.

Now, I have a data set where each row is

a sentence of arbitrary length.

And I've decided I want to train a windowed model to predict

a new word, given some words to start.

So I need to take these data set rows and make

THEM all equal length.

Once again, without leaving TensorFlow data sets,

I can split, flatten, and regroup my rows

so that each row is a set of 11 words.

But I want that last word in each row

to be a separate label.

So I just define a simple row-wise function

or pop out my labels.

And voila, we have pairs of examples and labels.

But there's still a problem with this data.

Why can't I just feed this into a dense network

or any arbitrary model?

Well, it's words.

Machine learning models don't speak English.

And, of course, the problem with this data set

is that we need numbers--

only numbers, not lines, just numbers.

So we're going to need to transform our input sentences

into numeric representations.

And the way that we do this is with preprocessing

layers, which are highly exciting

and a new addition to TensorFlow 2.

So preprocessing layers were recently

reviewed as part of the Keras API specification.

And there are a set of layers that

take care of data transformations

that are outside the training path.

These follow the same APIs as normal layers

and can be called and composed just like layers.

They play nicely with TensorFlow data sets as well.

So you can paralyze preprocessing transformations.

And importantly, just like normal Keras layers,

these become part of your model and therefore, get

serialized in the SavedModel and become part of inference

to prediction, which is critical to minimizing training

serving SKU.

This is the set of preprocessing layers

that is already complete.

They are experimental in 2.2, as we ensure that the APIs match

how you use them.

So please check them out for all of your preprocessing needs.

So now we know how to get our data into the correct format.

And the next step is to build a language model that

can learn a representation of all of these words

so that we can generate text on the fly.

One of the classic models used for text translation

and generation is a sequence-to-sequence model.

A sequence-to-sequence model typically has two parts--

an encoder that uses RNN blocks to encode

input data and a decoder that borrows state from the encoder

in order to correctly predict the target outputs.

Now, sequence-to-sequence models have some complicated

moving parts.

It's not a simple feed-forward network,

and there are a lot of parameters to keep track of.

But luckily, we don't have to go it alone here.

TensorFlow AddOns is a community-maintained repository

built on top of TensorFlow 2 that

provides especially complex or experimental layers,

optimizers, and other utilities.

And one such utility is the seq2seq package,

which provides a number of layers and classes

that make building sequence-to-sequence models

much easier.

And you'll see me use these throughout this example.

Because the architecture of the sequence-to-sequence model

is fairly complex and requires special state passing,

we're going to subclass the Keras-based model

and build our network explicitly.

We start here with the text factorization layer

we've already discussed since that's

going to be what we feed in our input data

through to convert it to indices.

After we vectorize the inputs, we're

going to pass them through the encoder blocks--

first in embedding, then in LSTM.

And you'll note that in our init function here,

we're just configuring the layers.

And we're not actually passing any data through them yet.

In Keras, each layer has two stages--

the construction, when you parametrize

your layer, as seen here, and the calling

of the layer, which will come later

when we pass our data through.

The decoder is somewhat more complicated.

But here, we can leverage the TensorFlow AddOn sampler

and decoder and set up the decoder LSTM

and connect it to projection layer, which

is our final dense layer that maps to the vocabulary

we want to predict, plus two tokens for predicting calls

and out of vocabulary words.

The final set of layers we will set up

are a pair of attention layers.

There is a large and growing field of attention, research,

and machine learning.

But given the time constraints of this talk,

I'm only going to give you a very hand wavy explanation

and say, attention is a technique

that allows the model to track intermediate states as it

steps long sequences.

And it will allow the model to give more weight

to certain time steps of a sequence

when predicting the final word.

Here, we use a simple dense attention

layer that comes with tf.keras.

And you'll see how we connect this between our encoder

and decoder in a minute.

So we have a lot of state to pass between our encoder

and decoder, which means we can't just use

the standard Keras fit call.

And one of the things I'm most excited about in TensorFlow 2

is that we have refactored and cleaned

up the Keras training loop.

And we've made it much more modular.

So instead of overriding the entire fit loop

or throwing it out altogether, I'm

just going to define a single forward pass

from my model in this special function,

train_step, and that is going to get called

by model.fit with each step.

So we overwrite train_step in our encoder decoder model.

That train_step is going to get one batch of data at a time,

so we just need to define the forward pass for that one

batch.

And the first thing we do here is unpack our data

and separate the example from the label.

You'll note that we call our own vectorization layer here

to ensure that our input strings get correctly

transformed to indices.

Next, still inside our single training step,

we're going to record our forward pass

under a gradient take.

Anything that needs to be back propped through

should go under here.

So while the vectorized layer or preprocessing layer was out

of the tape, our encoder embedding, LSTM, and so forth

all belong under the tape.

And here we pass our inputs through the set

of layers we defined on our init to encode them.

We also set up our attention layers

to track the intermediate state coming out

of the encoding layers.

And next, we decode, which is to say

we try to predict our targets using

the state from our encoder.

The decoder here, we'll go over the many epics it runs for,

train its own weights separate from the encoder weights.

And in concert, the encoder and decoder

will learn to predict text based on the outputs.

And now that we've run all of our layers

necessary to form the forward pass, we can compute the loss

and collect the outputs of this step

so that they can be optimized.

The Keras model takes care of collecting variables

and setting the optimizer, so we just

choose how we want to pass things through here.

As the final step, we collect and return

the metrics we set in our model.

The next step is to pick an optimizer loss and accuracy

metric for our model.

And these are going to govern the actual training

and optimization process.

And we select them from a bunch of independently parametrizable