We are thrilled to introduce keras3, the next version of the Keras R package. keras3 is a ground-up rebuild of {keras}, maintaining the beloved features of the original while refining and simplifying the API based on valuable insights gathered over the past few years. Keras provides a complete toolkit for building deep learning models in R—it’s never been easier to build, train, evaluate, and deploy deep learning models.
Installation
To install Keras 3:
install.packages("keras3")
library(keras3)
install_keras()
What’s new: Documentation
Great documentation is essential, and we’ve worked hard to make sure that keras3 has excellent documentation, both now, and in the future. Keras 3 comes with a full refresh of the website: https://keras.posit.co. There, you will find guides, tutorials, reference pages with rendered examples, and a new examples gallery. All the reference pages and guides are also available via R’s built-in help system.
In a fast moving ecosystem like deep learning, creating great documentation and wrappers once is not enough. There also need to be workflows that ensure the documentation is up-to-date with upstream dependencies. To accomplish this, {keras3} includes two new maintainer features that ensure the R documentation and function wrappers will stay up-to-date:
- We now take snapshots of the upstream documentation and API surface. With each release, all R documentation is rebased on upstream updates. This workflow ensures that all R documentation (guides, examples, vignettes, and reference pages) and R function signatures stay up-to-date with upstream. This snapshot-and-rebase functionality is implemented in a new standalone R package,
{doctether}, which may be useful for R package maintainers needing to keep documentation in parity with dependencies. - All examples and vignettes can now be evaluated and rendered during a package build. This ensures that no stale or broken example code makes it into a release. It also means all user facing example code now additionally serves as an extended suite of snapshot unit and integration tests. Evaluating code in vignettes and examples is still not permitted according to CRAN restrictions. We work around the CRAN restriction by adding additional package build steps that pre-render examples and vignettes.
Multi-backend support
Soon after its launch in 2015, Keras featured support for most popular deep learning frameworks: TensorFlow, Theano, MXNet, and CNTK. Over time, the landscape shifted; Theano, MXNet, and CNTK were retired, and TensorFlow surged in popularity. In 2021, three years ago, TensorFlow became the premier and only supported Keras backend. Now, the landscape has shifted again. Keras 3 brings the return of multi-backend support.
Choose a backend by calling:
use_backend("jax") # or “tensorflow”, “torch”, “numpy”
The default backend continues to be TensorFlow, which is the best choice for most users today; for small-to-medium sized models this is still the fastest backend. However, each backend has different strengths, and being able to switch easily will let you adapt to changes as your project, or the frameworks themselves, evolve. Today, switching to the Jax backend can, for some model types, bring substantial speed improvements. Jax is also the only backend that has support for a new model parallelism distributed training API. Switching to Torch can be helpful during development, often producing simpler trackbacks while debugging.
Keras 3 also lets you incorporate any pre-existing Torch, Jax, or Flax module as a standard Keras layer by using the appropriate wrapper, letting you build atop existing projects with Keras. For example, train a Torch model using the Keras high-level training API (compile() + fit()), or include a Flax module as a component of a larger Keras model. The new multi-backend support lets you use Keras à la carte.
The ‘Ops’ family
{keras3} introduces a new “Operations” family of function. The Ops family, currently with over 200 functions, provides a comprehensive suite of operations typically needed when operating on nd-arrays for deep learning. The Operation family supersedes and greatly expands on the former family of backend functions prefixed with k_ in the {keras} package.
The Ops functions let you write backend-agnostic code. They provide a uniform API, regardless of if you’re working with TensorFlow Tensors, Jax Arrays, Torch Tensors, Keras Symbolic Tensors, NumPy arrays, or R arrays. The Ops functions:
- all start with prefix
op_(e.g.,op_stack()) - all are pure functions (they produce no side-effects)
- all use consistent 1-based indexing, and coerce doubles to integers as needed
- all are safe to use with any backend (
tensorflow,jax,torch,numpy) - all are safe to use in both eager and graph/jit/tracing modes
The Ops API includes:
- The entirety of the
NumPyAPI (numpy.*) - The
TensorFlowNN API (tf.nn.*) - Common linear algebra functions (A subset of
scipy.linalg.*) - A subfamily of image transformers
- A comprehensive set of loss functions
- And more!
Ingest tabular data with layer_feature_space()
keras3 provides a new set of functions for building models that ingest tabular data: layer_feature_space() and a family of feature transformer functions (prefix, feature_) for building keras models that can work with tabular data, either as inputs to a keras model, or as preprocessing steps in a data loading pipeline (e.g., a tfdatasets::dataset_map()). See the reference page and an example usage in a full end-to-end example to learn more.
New Subclassing API
The subclassing API has been refined and extended to more Keras types. Define subclasses simply by calling:
Layer(), Loss(), Metric(), Callback(), Constraint(), Model(), and LearningRateSchedule(). Defining {R6} proxy classes is no longer necessary. Additionally the documentation page for each of the subclassing functions now contains a comprehensive listing of all the available attributes and methods for that type. Check out ?Layer to see what’s possible.
Saving and Export
Keras 3 brings a new model serialization and export API. It is now much simpler to save and restore models, and also, to export them for serving.
save_model()/load_model(): A new high-level file format (extension: .keras) for saving and restoring a full model. The file format is backend-agnostic. This means that you can convert trained models between backends, simply by saving with one backend, and then loading with another. For example, train a model usingJax, and then convert toTensorflowfor export.export_savedmodel(): Export just the forward pass of a model as a compiled artifact for inference withTF Servingor (soon)Posit Connect. This is the easiest way to deploy aKerasmodel for efficient and concurrent inference serving, all without any R or Python runtime dependency.- Lower level entry points:
save_model_weights()/load_model_weights(): save just the weights as .h5 files.save_model_config()/load_model_config(): save just the model architecture as a json file.register_keras_serializable(): Register custom objects to enable them to be serialized and deserialized.serialize_keras_object()/deserialize_keras_object(): Convert any Keras object to an R list of simple types that is safe to convert to JSON or rds. See the new Serialization and Saving vignette for more details and examples.
New random family
A new family of random tensor generators. Like the Ops family, these work with all backends. Additionally, all the RNG-using methods have support for stateless usage when you pass in a seed generator. This enables tracing and compilation by frameworks that have special support for stateless, pure, functions, like Jax. See ?random_seed_generator() for example usage.
Other additions:
- New
shape()function, one-stop utility for working with tensor shapes in all contexts. - New and improved
print(model)andplot(model)method. See some examples of output in the Functional API guide - All new
fit()progress bar and live metrics viewer output, including new dark-mode support in the RStudio IDE. - New
configfamily, a curated set of functions for getting and settingKerasglobal configurations.
Migrating from {keras} to {keras3}
{keras3} is ultimately a preview of the future Keras package. If you’re writing new code today, you can start using {keras3} right away. If you have legacy code that uses {keras}, you are encouraged to update the code for {keras3}. For many high-level API functions, such as layer_dense(), fit(), and keras_model(), minimal to no changes are required. However there is a long tail of small changes that you might need to make when updating code that made use of the lower-level Keras API. Some of those are documented here: https://keras.io/guides/migrating_to_keras_3/. If you’re running into issues or have questions about updating, don’t hesitate to ask on https://github.com/rstudio/keras/issues or https://github.com/rstudio/keras/discussions. The {keras} and {keras3} packages will coexist while the community transitions. During…
Source link
