New: OpenVINO backend

OpenVINO is now available as an infererence-only Keras backend. You can start using it by setting the backend field to "openvino" in your keras.json config file.

OpenVINO is a deep learning inference-only framework tailored for CPU (x86, ARM), certain GPUs (OpenCL capable, integrated and discrete) and certain AI accelerators (Intel NPU).

Because OpenVINO does not support gradients, you cannot use it for training (e.g. model.fit()) -- only inference. You can train your models with the JAX/TensorFlow/PyTorch backends, and when trained, reload them with the OpenVINO backend for inference on a target device supported by OpenVINO.

New: ONNX model export

You can now export your Keras models to the ONNX format from the JAX, TensorFlow, and PyTorch backends.

Just pass format="onnx" in your model.export() call:

# Export the model as a ONNX artifact
model.export("path/to/location", format="onnx")

# Load the artifact in a different process/environment
ort_session = onnxruntime.InferenceSession("path/to/location")

# Run inference
ort_inputs = {
    k.name: v for k, v in zip(ort_session.get_inputs(), input_data)
}
predictions = ort_session.run(None, ort_inputs)

New: Scikit-Learn API compatibility interface

It's now possible to easily integrate Keras models into Sciki-Learn pipelines! The following wrapper classes are available:

• keras.wrappers.SKLearnClassifier: implements the sklearn Classifier API
• keras.wrappers.SKLearnRegressor: implements the sklearn Regressor API
• keras.wrappers.SKLearnTransformer: implements the sklearn Transformer API

Other feature additions

• Add new ops:
  • Add keras.ops.diagflat
  • Add keras.ops.unravel_index
• Add new activations:
  • Add sparse_plus activation
  • Add sparsemax activation
• Add new image augmentation and preprocessing layers:
  • Add keras.layers.RandAugment
  • Add keras.layers.Equalization
  • Add keras.layers.MixUp
  • Add keras.layers.RandomHue
  • Add keras.layers.RandomGrayscale
  • Add keras.layers.RandomSaturation
  • Add keras.layers.RandomColorJitter
  • Add keras.layers.RandomColorDegeneration
  • Add keras.layers.RandomSharpness
  • Add keras.layers.RandomShear
• Add argument axis to tversky loss

JAX specific changes

• Add support for JAX named scope

TensorFlow specific changes

• Make keras.random.shuffle XLA compilable

PyTorch specific changes

• Add support for model.export() and keras.export.ExportArchive with the PyTorch backend, supporting both the TF SavedModel format and the ONNX format.

New Contributors

• @LavanyaKV1234 made their first contribution in #20553
• @jakubxy08 made their first contribution in #20563
• @dhantule made their first contribution in #20565
• @roebel made their first contribution in #20575
• @Surya2k1 made their first contribution in #20613
• @edge7 made their first contribution in #20584
• @adrinjalali made their first contribution in #20599
• @mmicu made their first contribution in #20655
• @rkazants made their first contribution in #19727
• @lkk7 made their first contribution in #20682
• @Furkan-rgb made their first contribution in #20684
• @punkeel made their first contribution in #20694
• @kas2020-commits made their first contribution in #20709

Full Changelog: v3.7.0...v3.8.0

API changes

• Add flash_attention argument to keras.ops.dot_product_attention and to keras.layers.MultiHeadAttention.
• Add keras.layers.STFTSpectrogram layer (to extract STFT spectrograms from inputs as a preprocessing step) as well as its initializer keras.initializers.STFTInitializer.
• Add celu, glu, log_sigmoid, hard_tanh, hard_shrink, squareplus activations.
• Add keras.losses.Circle loss.
• Add image visualization utilities keras.visualization.draw_bounding_boxes, keras.visualization.draw_segmentation_masks, keras.visualization.plot_image_gallery, keras.visualization.plot_segmentation_mask_gallery.
• Add double_checkpoint argument to BackupAndRestore to save a fallback checkpoint in case the first checkpoint gets corrupted.
• Add bounding box preprocessing support to image augmentation layers CenterCrop, RandomFlip, RandomZoom, RandomTranslation, RandomCrop.
• Add keras.ops.exp2, keras.ops.inner operations.

Performance improvements

• JAX backend: add native Flash Attention support for GPU (via cuDNN) and TPU (via a Pallas kernel). Flash Attention is now used automatically when the hardware supports it.
• PyTorch backend: add native Flash Attention support for GPU (via cuDNN). It is currently opt-in.
• TensorFlow backend: enable more kernel fusion via bias_add.
• PyTorch backend: add support for Intel XPU devices.

New Contributors

• @mostafa-mahmoud made their first contribution in #20313
• @TrAyZeN made their first contribution in #20321
• @dryglicki made their first contribution in #20353
• @jm-willy made their first contribution in #20352
• @Gopi-Uppari made their first contribution in #20377
• @nicolaspi made their first contribution in #20383
• @sineeli made their first contribution in #20368
• @LakshmiKalaKadali made their first contribution in #20403
• @mwtoews made their first contribution in #20427
• @mrry made their first contribution in #20438
• @rohithpudari made their first contribution in #20447
• @ma7555 made their first contribution in #20452
• @jakevdp made their first contribution in #20469
• @lcs-crr made their first contribution in #20503
• @rameshdange5191 made their first contribution in #20525

Full Changelog: v3.6.0...v3.7.0

Highlights

• New file editor utility: keras.saving.KerasFileEditor. Use it to inspect, diff, modify and resave Keras weights files. See basic workflow here.
• New keras.utils.Config class for managing experiment config parameters.

BREAKING changes

• When using keras.utils.get_file, with extract=True or untar=True, the return value will be the path of the extracted directory, rather than the path of the archive.

Other changes and additions

• Logging is now asynchronous in fit(), evaluate(), predict(). This enables 100% compact stacking of train_step calls on accelerators (e.g. when running small models on TPU).
  • If you are using custom callbacks that rely on on_batch_end, this will disable async logging. You can force it back by adding self.async_safe = True to your callbacks. Note that the TensorBoard callback isn't considered async safe by default. Default callbacks like the progress bar are async safe.
• Added keras.saving.KerasFileEditor utility to inspect, diff, modify and resave Keras weights file.
• Added keras.utils.Config class. It behaves like a dictionary, with a few nice features:
  • All entries are accessible and settable as attributes, in addition to dict-style (e.g. config.foo = 2 or config["foo"] are both valid)
  • You can easily serialize it to JSON via config.to_json().
  • You can easily freeze it, preventing future changes, via config.freeze().
• Added bitwise numpy ops:
  • bitwise_and
  • bitwise_invert
  • bitwise_left_shift
  • bitwise_not
  • bitwise_or
  • bitwise_right_shift
  • bitwise_xor
• Added math op keras.ops.logdet.
• Added numpy op keras.ops.trunc.
• Added keras.ops.dot_product_attention.
• Added keras.ops.histogram.
• Allow infinite PyDataset instances to use multithreading.
• Added argument verbose in keras.saving.ExportArchive.write_out() method for exporting TF SavedModel.
• Added epsilon argument in keras.ops.normalize.
• Added Model.get_state_tree() method for retrieving a nested dict mapping variable paths to variable values (either as numpy arrays or backend tensors (default)). This is useful for rolling out custom JAX training loops.
• Added image augmentation/preprocessing layers keras.layers.AutoContrast, keras.layers.Solarization.
• Added keras.layers.Pipeline class, to apply a sequence of layers to an input. This class is useful to build a preprocessing pipeline. Compared to a Sequential model, Pipeline features a few important differences:
  • It's not a Model, just a plain layer.
  • When the layers in the pipeline are compatible with tf.data, the pipeline will also remain tf.data compatible, independently of the backend you use.

New Contributors

• @alexhartl made their first contribution in #20125
• @Doch88 made their first contribution in #20156
• @edbosne made their first contribution in #20151
• @ghsanti made their first contribution in #20185
• @joehiggi1758 made their first contribution in #20223
• @AryazE made their first contribution in #20228
• @sanskarmodi8 made their first contribution in #20237
• @himalayo made their first contribution in #20262
• @nate2s made their first contribution in #20305
• @DavidLandup0 made their first contribution in #20316

Full Changelog: v3.5.0...v3.6.0

What's Changed

• Add integration with the Hugging Face Hub. You can now save models to Hugging Face Hub directly from keras.Model.save() and load .keras models directly from Hugging Face Hub with keras.saving.load_model().
• Ensure compatibility with NumPy 2.0.
• Add keras.optimizers.Lamb optimizer.
• Improve keras.distribution API support for very large models.
• Add keras.ops.associative_scan op.
• Add keras.ops.searchsorted op.
• Add keras.utils.PyDataset.on_epoch_begin() method.
• Add data_format argument to keras.layers.ZeroPadding1D layer.
• Bug fixes and performance improvements.

Full Changelog: v3.4.1...v3.5.0

This is a minor bugfix release.

Highlights

• Add support for arbitrary, deeply nested input/output structures in Functional models (e.g. dicts of dicts of lists of inputs or outputs...)
• Add support for optional Functional inputs.
• Introduce keras.dtype_policies.DTypePolicyMap for easy configuration of dtype policies of nested sublayers of a subclassed layer/model.
• New ops:
  • keras.ops.argpartition
  • keras.ops.scan
  • keras.ops.lstsq
  • keras.ops.switch
  • keras.ops.dtype
  • keras.ops.map
  • keras.ops.image.rgb_to_hsv
  • keras.ops.image.hsv_to_rgb

What's changed

• Add support for float8 inference for Dense and EinsumDense layers.
• Add custom name argument in all Keras Applications models.
• Add axis argument in keras.losses.Dice.
• Enable keras.utils.FeatureSpace to be used in a tf.data pipeline even when the backend isn't TensorFlow.
• StringLookup layer can now take tf.SparseTensor as input.
• Metric.variables is now recursive.
• Add training argument to Model.compute_loss().
• Add dtype argument to all losses.
• keras.utils.split_dataset now supports nested structures in dataset.
• Bugs fixes and performance improvements.

Full Changelog: v3.3.3...v3.4.0

This is a minor bugfix release.

This is a simple fix release that re-surfaces legacy Keras 2 APIs that aren't part of Keras package proper, but that are still featured in tf.keras. No other content has changed.

This is a simple fix release that moves the legacy _tf_keras API directory to the root of the Keras pip package. This is done in order to preserve import paths like from tensorflow.keras import layers without making any changes to the TensorFlow API files.

No other content has changed.

What's Changed

• Introduce float8 training.
• Add LoRA to ConvND layers.
• Add keras.ops.ctc_decode for JAX and TensorFlow.
• Add keras.ops.vectorize, keras.ops.select.
• Add keras.ops.image.rgb_to_grayscale.
• Add keras.losses.Tversky loss.
• Add full bincount and digitize sparse support.
• Models and layers now return owned metrics recursively.
• Add pickling support for Keras models. Note that pickling is not recommended, prefer using Keras saving APIs.
• Bug fixes and performance improvements.

In addition, the codebase structure has evolved:

• All source files are now in keras/src/.
• All API files are now in keras/api/.
• The codebase structure stays unchanged when building the Keras pip package. This means you can pip install Keras directly from the GitHub sources.

New Contributors

• @kapoor1992 made their first contribution in #19484
• @IMvision12 made their first contribution in #19393
• @alanwilter made their first contribution in #19438
• @chococigar made their first contribution in #19323
• @LukeWood made their first contribution in #19555
• @AlexanderLavelle made their first contribution in #19575

Full Changelog: v3.2.1...v3.3.0