We’re thrilled to introduce keras3, the subsequent model of the Keras R
bundle. keras3 is a ground-up rebuild of {keras}, sustaining the
beloved options of the unique whereas refining and simplifying the API
primarily based on beneficial insights gathered over the previous few years.

Keras gives a whole toolkit for constructing deep studying fashions in
R—it’s by no means been simpler to construct, prepare, consider, and deploy deep
studying fashions.

Set up

To put in Keras 3:

set up.packages("keras3")
library(keras3)
install_keras()

Keras 3 comes with a full refresh of the web site:
https://keras.posit.co. There, you can find guides, tutorials,
reference pages with rendered examples, and a brand new examples gallery. All
the reference pages and guides are additionally accessible by way of R’s built-in assist
system.

In a fast paced ecosystem like deep studying, creating nice
documentation and wrappers as soon as shouldn’t be sufficient. There additionally must be
workflows that make sure the documentation is up-to-date with upstream
dependencies. To perform this, {keras3} consists of two new maintainer
options that make sure the R documentation and performance wrappers will keep
up-to-date:

• We now take snapshots of the upstream documentation and API floor.
  With every launch, all R documentation is rebased on upstream
  updates. This workflow ensures that every one R documentation (guides,
  examples, vignettes, and reference pages) and R perform signatures
  keep up-to-date with upstream. This snapshot-and-rebase
  performance is applied in a brand new standalone R bundle,
  {doctether}, which can
  be helpful for R bundle maintainers needing to maintain documentation in
  parity with dependencies.

• All examples and vignettes can now be evaluated and rendered throughout
  a bundle construct. This ensures that no stale or damaged instance code
  makes it right into a launch. It additionally means all person dealing with instance code
  now moreover serves as an prolonged suite of snapshot unit and
  integration assessments.

  Evaluating code in vignettes and examples continues to be not permitted
  in keeping with CRAN restrictions. We work across the CRAN restriction
  by including further bundle construct steps that pre-render
  examples
  and
  vignettes.

Mixed, these two options will make it considerably simpler for Keras
in R to take care of characteristic parity and up-to-date documentation with the
Python API to Keras.

Multi-backend assist

Quickly after its launch in 2015, Keras featured assist for hottest
deep studying frameworks: TensorFlow, Theano, MXNet, and CNTK. Over
time, the panorama shifted; Theano, MXNet, and CNTK had been retired, and
TensorFlow surged in recognition. In 2021, three years in the past, TensorFlow
grew to become the premier and solely supported Keras backend. Now, the panorama
has shifted once more.

Keras 3 brings the return of multi-backend assist. Select a backend by
calling:

use_backend("jax") # or "tensorflow", "torch", "numpy"

{keras3} introduces a brand new “Operations” household of perform. The Ops
household, presently with over 200
capabilities,
gives a complete suite of operations sometimes wanted when
working on nd-arrays for deep studying. The Operation household
supersedes and significantly expands on the previous household of backend capabilities
prefixed with k_ within the {keras} bundle.

The Ops capabilities allow you to write backend-agnostic code. They supply a
uniform API, no matter in the event you’re working with TensorFlow Tensors,
Jax Arrays, Torch Tensors, Keras Symbolic Tensors, NumPy arrays, or R
arrays.

The Ops capabilities:

• all begin with prefix op_ (e.g., op_stack())
• all are pure capabilities (they produce no side-effects)
• all use constant 1-based indexing, and coerce doubles to integers
  as wanted
• all are protected to make use of with any backend (tensorflow, jax, torch, numpy)
• all are protected to make use of in each keen and graph/jit/tracing modes

The Ops API consists of:

• The whole thing of the NumPy API (numpy.*)
• The TensorFlow NN API (tf.nn.*)
• Widespread linear algebra capabilities (A subset of scipy.linalg.*)
• A subfamily of picture transformers
• A complete set of loss capabilities
• And extra!

Ingest tabular information with layer_feature_space()

keras3 gives a brand new set of capabilities for constructing fashions that ingest
tabular information: layer_feature_space() and a household of characteristic
transformer capabilities (prefix, feature_) for constructing keras fashions
that may work with tabular information, both as inputs to a keras mannequin, or
as preprocessing steps in a knowledge loading pipeline (e.g., a
tfdatasets::dataset_map()).

See the reference
web page and an
instance utilization in a full end-to-end
instance
to be taught extra.

New Subclassing API

The subclassing API has been refined and prolonged to extra Keras
sorts.
Outline subclasses just by calling: Layer(), Loss(), Metric(),
Callback(), Constraint(), Mannequin(), and LearningRateSchedule().
Defining {R6} proxy lessons is not crucial.

Moreover the documentation web page for every of the subclassing
capabilities now accommodates a complete itemizing of all of the accessible
attributes and strategies for that kind. Try
?Layer to see what’s
potential.

Saving and Export

Keras 3 brings a brand new mannequin serialization and export API. It’s now a lot
easier to save lots of and restore fashions, and in addition, to export them for
serving.

• save_model()/load_model():
  A brand new high-level file format (extension: .keras) for saving and
  restoring a full mannequin.

  The file format is backend-agnostic. This implies that you could convert
  educated fashions between backends, just by saving with one backend,
  after which loading with one other. For instance, prepare a mannequin utilizing Jax,
  after which convert to Tensorflow for export.

• export_savedmodel():
  Export simply the ahead move of a mannequin as a compiled artifact for
  inference with TF
  Serving or (quickly)
  Posit Join. This
  is the best solution to deploy a Keras mannequin for environment friendly and
  concurrent inference serving, all with none R or Python runtime
  dependency.

• Decrease stage entry factors:

  • save_model_weights() / load_model_weights():
    save simply the weights as .h5 recordsdata.
  • save_model_config() / load_model_config():
    save simply the mannequin structure as a json file.

• register_keras_serializable():
  Register customized objects to allow them to be serialized and
  deserialized.

• serialize_keras_object() / deserialize_keras_object():
  Convert any Keras object to an R listing of straightforward sorts that’s protected
  to transform to JSON or rds.

• See the brand new Serialization and Saving
  vignette
  for extra particulars and examples.

New random household

A brand new household of random tensor
mills.
Just like the Ops household, these work with all backends. Moreover, all of the
RNG-using strategies have assist for stateless utilization whenever you move in a
seed generator. This allows tracing and compilation by frameworks that
have particular assist for stateless, pure, capabilities, like Jax. See
?random_seed_generator()
for instance utilization.

Different additions:

• New form()
  perform, one-stop utility for working with tensor shapes in all
  contexts.

• New and improved print(mannequin) and plot(mannequin) technique. See some
  examples of output within the Useful API
  information

• All new match() progress bar and dwell metrics viewer output,
  together with new dark-mode assist within the RStudio IDE.

• New config
  household,
  a curated set of capabilities for getting and setting Keras international
  configurations.

• All the different perform households have expanded with new members:

Migrating from {keras} to {keras3}

{keras3} supersedes the {keras} bundle.

In the event you’re writing new code right this moment, you can begin utilizing {keras3} proper
away.

When you have legacy code that makes use of {keras}, you’re inspired to
replace the code for {keras3}. For a lot of high-level API capabilities, such
as layer_dense(), match(), and keras_model(), minimal to no adjustments
are required. Nevertheless there’s a lengthy tail of small adjustments that you simply
may have to make when updating code that made use of the lower-level
Keras API. A few of these are documented right here:
https://keras.io/guides/migrating_to_keras_3/.

In the event you’re operating into points or have questions on updating, don’t
hesitate to ask on https://github.com/rstudio/keras/points or
https://github.com/rstudio/keras/discussions.

The {keras} and {keras3} packages will coexist whereas the neighborhood
transitions. Throughout the transition, {keras} will proceed to obtain
patch updates for compatibility with Keras v2, which continues to be
printed to PyPi below the bundle identify tf-keras. After tf-keras is
not maintained, the {keras} bundle can be archived.

Abstract

In abstract, {keras3} is a sturdy replace to the Keras R bundle,
incorporating new options whereas preserving the convenience of use and
performance of the unique. The brand new multi-backend assist,
complete suite of Ops capabilities, refined mannequin serialization API,
and up to date documentation workflows allow customers to simply take
benefit of the most recent developments within the deep studying neighborhood.

Whether or not you’re a seasoned Keras person or simply beginning your deep
studying journey, Keras 3 gives the instruments and suppleness to construct,
prepare, and deploy fashions with ease and confidence. As we transition from
Keras 2 to Keras 3, we’re dedicated to supporting the neighborhood and
guaranteeing a easy migration. We invite you to discover the brand new options,
try the up to date documentation, and be part of the dialog on our
GitHub discussions web page. Welcome to the subsequent chapter of deep studying in
R with Keras 3!