Transcription of: How to Import a Keras model into TensorFlow.js



YUFENG GUO:

TensorFlow.js is awesome. Keras is awesome. Wouldn't it be extra amazing

if they could work together? Well, I got great news for you. They do! Stay tuned to find out just how

to import your trained Keras model into TensorFlow.js. Welcome to AI Adventures, where

we explore the art, science, and tools of machine learning. My name is Yufeng Guo,

and on this episode, we're going to look at what

it takes to import a Keras model into TensorFlow.js. TensorFlow.js has

many functionalities, but perhaps one of

the most exciting ones is its ability to operate in

collaboration with other tools. We're really seeing

a future where you can do machine learning

across a range of platforms, languages, and

devices, all optimized to perform at their

peak performance in the situation they're

most well-suited for. The tricky thing about loading

a Keras model into TensorFlow.js is that Keras models are

typically saved as HDF5 files, using something like

model.save('my_model.h5'). TensorFlow.js provides a tool,

the TensorFlow.js converter, which is used to convert your

Keras model into a format that TensorFlow.js

can consume, which is a folder containing a

file called model.json. Let's see that in context

with the Keras model that we've been using in

the last few episodes. In the episode where we

converted a Keras model to a TensorFlow estimator, we

also exported our Keras model as a .h5 file, so we'll

head on over to that Kernel and download the file

from the Output section. So I've got my Keras model

downloaded and named it keras_model.h5 just to keep

things clear which file is from which library. Now I need to run pip install

tensorflowjs in my Python environment to

install the converter. Notice that the pip library

name is tensorflowjs, even though the converter

we'll use is called tensorflowjs converter. OK, so we've got the converter

and we got the Keras model all in one place. One last step before we

run the converter, though. I'm going to create a

new folder because we need to store the outputs of

the TensorFlow.js model files. Let's just call it tfjs_files. Now we can run the

converter, passing in a flag to indicate that it

is a Keras model, and we'll also need to pass

the path to the Keras HDF5 file in that folder we just created

so it has a destination to put the new files in. Once our converter

has run, let's take a peek inside the folder. We see that it's got not only

the model.json file, but three additional files. There's a group1-shard1of1,

group2-shard1of1, and group3-shard1of1. These shards are here to

enable faster repeat loading by the browser. The files are below the

typical cache size limit, so they'll likely be

cached for subsequent calls when you're serving them. So after you've successfully

converted your Keras model to model.json, you've just

got one last step left-- got to load it into

your JavaScript. Thankfully, this

is a one-line task. Go ahead and use tf.modelload

to point to the URL where you have the files

posted, and TensorFlow.js takes care of the rest. It's worth pointing out

here that since the model is operating on the client

side, using JSON files, it's not a great use case if

you need to secure your model. For that, you'll have to

stick with client side models. All right. So now you're ready to

import your own Keras models into TensorFlow.js. Go forth and make the

next great machine learning-inspired

game on the web. Thanks for watching this

episode of "Cloud AI Adventures," and if you

enjoyed it, please like it and subscribe to get all

the latest episodes right when they come out. For now, try loading your

Keras model into TensorFlow.js and see just how easy it is.