Keras Core 0.1.7 release

Highlights

• Add ops keras.random.shuffle, keras.ops.image.map_coordinates
• Add support for tf.SparseTensor with TensorFlow backend (same level of support as legacy tf.keras)
• Make Discretization layer backend agnostic
• Bug fixes and performance improvements (in particular significantly reduced memory usage during training for JAX)

What's Changed

• Make jax2tf import conditional by @nkovela1 in #911
• Separate the metrics variables from non-trainable variables. by @qlzh727 in #910
• Update jax trainer function to save memory buffer. by @qlzh727 in #897
• [GHA] fix several codecov issues by @kiukchung in #912
• Add example to fine-tune TorchVision models using Keras by @soumik12345 in #909
• Refactor merging layers tests as parametrized tests. by @hertschuh in #913
• Add ops.random.shuffle by @james77777778 in #907
• Add ops.map_coordinates by @james77777778 in #906
• Makes ops.split in torch consistent with other backends by @james77777778 in #914
• Add: MultipleChoice with Transfer Learning example by @awsaf49 in #884
• Update test_case.py by @PatReis in #917
• Fix JAX RNN backend issue. by @qlzh727 in #924
• Adds JAX version checking for Export by @nkovela1 in #923
• make discretization backend agnostic by @divyashreepathihalli in #896
• Memory optimization for jax trainer. by @qlzh727 in #888
• [*.py] Standardise docstring usage of "Default to" by @SamuelMarks in #921
• Fix the inconsistency in affine_transform between numpy and jax by @james77777778 in #926
• Add support for Tensorflow SparseTensors: merging layers. by @hertschuh in #925
• fix torch error for bincount by @divyashreepathihalli in #927
• Refactor torch's affine_transform by @james77777778 in #929
• Increase test coverage + Fix save_model_to_hdf5 + Improve is_remote_path + Fix is_remote_path by @Faisal-Alsrheed in #900
• Slice serialization by @gleize in #932
• Some dtype fixes by @fchollet in #935
• Hacky fix for dictionary output with tf 2.14 by @mattdangerw in #933

New Contributors

• @awsaf49 made their first contribution in #884
• @PatReis made their first contribution in #917
• @SamuelMarks made their first contribution in #921

Full Changelog: v0.1.6...v0.1.7

Keras Core 0.1.6 release

Highlights

• Performance optimizations in normalization layers
• New op: ops.nn.moments
• Add TorchModuleWrapper to make it possible to use PyTorch Modules in Keras Core models with the PyTorch backend. Modules added as attributes of a layer are automatically wrapped (so that their weights are tracked).
• Add ignore_class argument in SparseCategoricalCrossentropy.
• Add new experimental data parallel and model parallel distribution API for JAX.

What's Changed

• Fix where op by @AakashKumarNain in #770
• Add skeleton for the common distribution backend. by @qlzh727 in #765
• Add more H5 backwards compatibility tests and lambda support by @nkovela1 in #775
• Add implementation for Distribution API by @qlzh727 in #776
• Removes tf.keras from transposed conv test by @sampathweb in #782
• Fix channels_first format bug for rescaling layer by @haifeng-jin in #792
• Fix argument order in stft and istft by @james77777778 in #790
• fixed squeeze axis inconsistency by @jackd in #788
• Replace tf.data with tf_data in preprocessing tests by @sampathweb in #787
• Reimplement the data parallel distribution. by @qlzh727 in #785
• Split fix by @jackd in #789
• Fix bug: Softmax function to support tuple axes + Add more tests by @Faisal-Alsrheed in #778
• Add LayoutMap which will be used for model parallelism by @qlzh727 in #793
• Fix bug: Log_Softmax with tuple axes + Add tests by @Faisal-Alsrheed in #794
• Port tab transformer example to keras core by @anas-rz in #621
• Update example losses on docstrings by @Frightera in #795
• Add model parallel distribution. by @qlzh727 in #797
• Increase-test-coverage-in-activations_test.py by @Faisal-Alsrheed in #799
• add ignore_class param to sparse CE loss by @AakashKumarNain in #599
• Convert pretraining_BERT.py example to keras core by @pranavvp16 in #801
• fixed apply_mask bug in losses by @jackd in #802
• Replace tf.keras in center-crop and resize tests by @sampathweb in #804
• Add a check for class_id argument by @Frightera in #796
• Add a dummy init for distribution, so it will be treat as a package. by @qlzh727 in #809
• Allow test using dynamic shapes to run on JAX. by @hertschuh in #810
• Keep "mse" as the metric name in the log by @haifeng-jin in #812
• Fixed misc mixed precision issues by @mattdangerw in #805
• Don't double cast layer norm weights in mixed precision by @mattdangerw in #815
• Add loss attribute for backward compatibility by @haifeng-jin in #813
• Add TorchModuleWrapper by @soumik12345 in #779
• Make sure the Test CI action fails on bad lint by @kiukchung in #818
• Format confusion_metrics_test by @mattdangerw in #821
• Fix the Jax distributed variable load issue. by @qlzh727 in #817
• Always save the build config when building a layer by @mattdangerw in #820
• Uses backend.shape in random preprocessing by @sampathweb in #822
• Don't lock functional models by @mattdangerw in #823
• Refactor conv_transpose padding by @guillaumebaquiast in #811
• Add rank>2 support for stft by @james77777778 in #825
• Fix reduce_shape + Add tests to operation_utils_test.py by @Faisal-Alsrheed in #826
• Fix weird broken test that never ran by @mattdangerw in #827
• Convert "English-to-Spanish translation with KerasNLP" to keras-core by @freedomtan in #777
• Replace assertTrue(isinstance(...)) with assertIsInstance for easier … by @hertschuh in #831
• Improve the flexibility of standardize_dtype and fix pad in torch backend by @james77777778 in #828
• Fix the distribution logic for the JAX trainer. by @qlzh727 in #832
• Add a docstring for compile by @mattdangerw in #838
• Enable codecov by @kiukchung in #840
• Add support for Tensorflow SparseTensors: core classes. by @hertschuh in #839
• bert-mlm port to keras-core with tf backend by @Mrutyunjay01 in #843
• Enable codecov flags and coverage carryover for keras_core and keras_… by @kiukchung in #841
• Update minimum python version to 3.9 to align with TF/JAX by @sampathweb in #848
• Update the eager_build logic for jax trainer. by @qlzh727 in #845
• Update return type for optimizer.apply by @mattdangerw in #850
• Add test,codecov,pypi status badges to README by @kiukchung in #852
• Fix autocast scope again by @mattdangerw in #849
• Add Export for JAX Backend by @nkovela1 in #819
• Increase test coverage + Fixing bug in static_call by @Faisal-Alsrheed in #847
• Add a loss scale optimizer by @mattdangerw in #851
• Increase test coverage in keras_core/utils by @Faisal-Alsrheed in #856
• Increase test coverage in keras_core/trainers by @Faisal-Alsrheed in #860
• Add a note in README.md regarding pre-import of backend by @SuryanarayanaY in #863
• Add optimizer weights to the model summary. by @qlzh727 in #859
• Add support for Tensorflow SparseTensors: core layers. by @hertschuh in #854
• Increase test coverage in keras_core/backend by @Faisal-Alsrheed in #864
• Add support for Tensorflow SparseTensors: reshaping layers. by @hertschuh in #865
• Respect the $KERAS_HOME environment variable for get_file by @mattdangerw in #868
• Don't convert int array input to floatx by @mattdangerw in #867
• run_layer_test also runs symbolic calls when input_data is provided. by @hertschuh in #870
• Increase test coverage in utils + Fix Bug in cast_to_common_dtype by @Faisal-Alsrheed in #875
• Fix stop_gradient in np by @james77777778 in #872
• Add sparseness inference for TensorFlow backend ops and base layer. by @hertschuh in #878
• Fix summary when optimizer is None by @mattdangerw in #879
• Add ops.nn.moments and speed-up normalization layers by @james77777778 in #866
• Don't show optimizer params if not optimizer or not built by @mattdangerw in #885
• Export dtype utils in keras_core.backend by @mattdangerw in #886
• Set actions.yml to run with read-only permissions by @pnacht in #882
• Fix model.save API for legacy H5 format by @nkovela1 in #891
• Add missing dtypes. by @gleize in #887
• Increase tests in utils by @Faisal-Alsrheed in #880
• syntax fix in cct example by @anas-rz in #898
• Use ops.rsqrt, improve normalization layers and enable ops fusion in tflite by @james77777778 in #892
• fixes #550: move torch tensors to cpu before to numpy array by @dkgaraujo in #853

New Contributors

• @kiukchung made ...

Keras Core 0.1.5 release

Highlights

• Add model.export() flow.
• Add ops rsqrt, extract_sequences, rfft, stft, irfft, istft.
• Bug fixes and performance improvements.
• ops.shape now returns a tuple with all backends (rather than a TF tensor with the TF backend). The tuple may contain scalar tensors for dynamic dimensions, and contains integers otherwise for static dimensions.

What's Changed

• Ensure backwards compatibility for legacy H5 saving format by @nkovela1 in #682
• Encapsulate build logic by @grasskin in #679
• Fix the --install arg doesn't work by @haifeng-jin in #686
• Convert to tensor for jax ops.numpy.transpose by @mattdangerw in #687
• Fix custom functional reload issue by @fchollet in #693
• Update the softmax default to axis=-1 by @mattdangerw in #694
• Random Flip for batched inputs by @guillaumebaquiast in #698
• Fix TextVectorization + Sequential bug by @fchollet in #696
• Add Export for TF backend by @nkovela1 in #692
• Fix typo in WhileLoop docstring in core.py by @Faisal-Alsrheed in #703
• Add activation docstrings // ops.nn.py by @Frightera in #705
• Remove tf dependency from math_test by @Frightera in #706
• Little cleanup in backend - config.py / export.py by @Frightera in #709
• Fix Issue in SegmentMax's Symbolic Call in keras_core/ops/math.py by @Faisal-Alsrheed in #713
• Replace tf.keras for expected result in Conv1D Test by @sampathweb in #725
• Add some additional H5 saving test coverage by @nkovela1 in #724
• Always return a tuple from ops.shape by @mattdangerw in #722
• Only compute mean once in layernorm by @mattdangerw in #721
• Add rsqrt to ops API by @shivance in #708
• Fix position encoder in examples/.../token_learner.py by @guillaumebaquiast in #727
• Support keras.Input in get_source_inputs by @haifeng-jin in #732
• Convert cyclegan to keras-core by @freedomtan in #728
• Add rms_scaling in LayerNormalization by @shivance in #726
• Docstrings for numpy ops by @FayazRahman in #714
• Add extract_sequences, rfft and stft to ops.math by @james77777778 in #717
• Apply scipy LazyModule by @james77777778 in #734
• Support torch symbolic call for normalization layer in preprocessing by @haifeng-jin in #739
• Replace tf.keras with numpy for conv2d and conv3d tests by @sampathweb in #740
• Enhanced Docstrings and Examples in /ops/function.py, /ops/math.py and /ops/nn.py. by @Faisal-Alsrheed in #736
• Fixed uniqueness bug and bits->bytes by @jackd in #747
• Add Examples to numpy.py functions by @Faisal-Alsrheed in #745
• Use convert_to_numpy in hashing and tests by @james77777778 in #748
• Convert ddim by @freedomtan in #743
• Add irfft and istft by @james77777778 in #742
• ported hashed-crossing by @asingh9530 in #685
• Add a high level API for distribution for JAX backend. by @qlzh727 in #741
• An attempt to add MPS device to torch backend by @M7Saad in #750
• Remove tf.keras from depthwise and separable conv tests by @sampathweb in #752
• Fix bug + Add Tests + Enhance docstrings (shape_equal) by @Faisal-Alsrheed in #751
• Fix standardize_shape by @james77777778 in #753
• Fix naming in mobilenet_v3 and densenet by @james77777778 in #759
• Enhance docstrings + easy to understand examples (nn.py) by @Faisal-Alsrheed in #758
• Remove backend.DYNAMIC_SHAPES_OK as dynamic shapes are supported in a… by @hertschuh in #763

New Contributors

• @Faisal-Alsrheed made their first contribution in #703
• @shivance made their first contribution in #708
• @jackd made their first contribution in #747
• @M7Saad made their first contribution in #750

Full Changelog: v0.1.4...v0.1.5

Keras Core 0.1.4 release

Highlights

• Add Lion optimizer
• Enable model.predict() and model.evaluate() with the NumPy backend
• Add ops cosh, sinh, arctanh, arcsinh, arccosh
• Add ops segment_max, digitize
• Bug fixes and performance improvements

What's Changed

• Set min TF Version by @sampathweb in #619
• Avoid some tree operations by @haifeng-jin in #620
• Add Lion optimizer by @james77777778 in #610
• Typo in layer tutorial by @guillaumebaquiast in #624
• fix load of models with Add and Concat layers by @freedomtan in #626
• Port/test learnable resizer to keras core backend agnostic by @anas-rz in #622
• Port fixres keras core by @anas-rz in #630
• Only ignore xx_mask arguments for the build shapes dict by @mattdangerw in #632
• Add numpy trainer by @james77777778 in #633
• [split_dataset] migrating from tf.keras to keras_core by @asingh9530 in #505
• Add hyperbolic ops by @FayazRahman in #634
• Port visualizing what convnets learn by @anas-rz in #640
• Port the ImageClassifier guide to keras_core by @tirthasheshpatel in #608
• Port simsiam to keras-core by @anas-rz in #644
• Fix the ImageClassifier demo and minimize the use of tensorflow by @tirthasheshpatel in #645
• Port neural decision forest by @anas-rz in #631
• Removes ref in Global Average Pool and Elu/Prelu Activations by @sampathweb in #646
• Small fixes on image.extract_patches by @guillaumebaquiast in #647
• Fix the missing weight name for Dense layer by @qlzh727 in #648
• Update learning_rate type to float in learning_rate_scheduler's logs by @siyavash in #649
• Performance improvement for MHA by @haifeng-jin in #654
• Add digitize op for Discretization layer by @abuelnasr0 in #641
• support passing jax.random.PRNGKey inputs in jax by @GallagherCommaJack in #651
• Removes tf.keras from Pooling tests - Max, Average and Global pooling by @sampathweb in #657
• Add numpy bincount docstrings by @dranaivo in #655
• Add a segment_max op by @ianstenbit in #660
• Port mlp_image_classification.py to all backends by @guillaumebaquiast in #663
• update requirements for all cpu install by @haifeng-jin in #666
• Random rotation port by @kamathis4 in #584
• Port TensorFlow implementation of the RandomCrop layer by @soumik12345 in #493
• Backend agnostic port of object detection using vit by @soumik12345 in #668
• Remove self.support_jit = False in RandomFlip by @james77777778 in #665
• Port Compact Convolutional Transformers to backend agnostic by @guillaumebaquiast in #669
• Move mlp_image_classification to backend agnostic folder by @guillaumebaquiast in #671
• Change saving API routing to deprecate SavedModel format by @nkovela1 in #674
• Convert neural machine translation with transformer to keras-core by @freedomtan in #623
• Make ner transformers work again by @freedomtan in #675
• Do not default to file root by @grasskin in #678
• Convert "text generation with miniature gpt" to Keras Core by @freedomtan in #676
• torch image added by @asingh9530 in #659
• Added some docstrings ops/nn.py by @Frightera in #677

New Contributors

• @asingh9530 made their first contribution in #505
• @FayazRahman made their first contribution in #634
• @siyavash made their first contribution in #649
• @abuelnasr0 made their first contribution in #641
• @dranaivo made their first contribution in #655

Full Changelog: v0.1.3...v0.1.4

Keras Core 0.1.3 release

Highlights

• Add compatibility with legacy whole-model h5 saving and loading
• Add keras_core.ops.fft and keras_core.ops.fft2
• Add keras_core.ops.image.affine_transform
• Add keras_core.ops.image.extract_patches
• Bug fixes and performance improvements
• Now compatible with TensorFlow versions 2.8 to 2.13

What's Changed

• Add Nadam for torch backend by @haifeng-jin in #551
• Fix multi-dimension sample weights passed to mean metrics by @mattdangerw in #552
• Update operation_utils.py - Added docstrings by @sqali in #514
• Node test by @kamathis4 in #557
• Add affine_transform op to all backends by @james77777778 in #477
• Add FFT Ops by @abheesht17 in #480
• Added CategoryEncoding to keras_core.ops.nn by @hazemessamm in #490
• Symbolic Args test by @kamathis4 in #536
• Converted to Keras Core: DeepLabV3Plus by @anas-rz in #545
• test and move backend agnostic example by @anas-rz in #567
• Port few shot learning with reptile by @anas-rz in #564
• Add some additional TF SavedModel tests by @nkovela1 in #569
• Fixes Torch-GPU Test failure on initializers/random_initializers_test.py by @sampathweb in #575
• Keras tensor functions by @kamathis4 in #576
• add fori_loop op to all 3 backends by @GallagherCommaJack in #462
• Fix Module Utils _available attribute by @sampathweb in #579
• Add some docstrings to keras_core/ops/numpy.py by @guillaumebaquiast in #558
• Add random seed support for torch dropout by @haifeng-jin in #568
• Port Keypoint detection to backend agnostic-keras-core by @anas-rz in #546
• Make random flip backend agnostic by @AmedeoBiolatti in #515
• Convert RandomTranslation to the backend-agnostic implementation by @james77777778 in #572
• doc: DocString for numpy.amax by @pranavvp16 in #582
• Add backend-agnostic video transformers example by @soumik12345 in #583
• Converted cutmix example to Keras Core by @cosmo3769 in #527
• docString for numpy.amin by @pranavvp16 in #586
• Update Tests for specific backend imports by @sampathweb in #591
• implement patch extraction in ops by @anas-rz in #581
• Converted to keras core external attention by @anas-rz in #529
• Backend consistency in handling of empty tensor in min and max by @guillaumebaquiast in #592
• Fix more spots torch != numpy for reductions by @mattdangerw in #596
• Backend-agnostic port of Image classification with Vision Transformer by @soumik12345 in #589
• More backend agnostic in cutmix and mixup file by @cosmo3769 in #593
• Always install the correct version with pip_build by @mattdangerw in #595
• Export ops.cond by @AmedeoBiolatti in #577
• Convert pretrained word embeddings to Keras Core by @freedomtan in #600
• port gcam backend agnostic by @anas-rz in #601
• Converted NER with transformers example to keras_core by @pranavvp16 in #594
• Add tf.keras backwards compat for nearly all non-experimental symbols by @fchollet in #603
• Cache self.call() signature by @haifeng-jin in #606
• Porting over changes to Torch and JAX keras-core distribution guides … by @hertschuh in #607
• Improve TensorFlowTrainer compatibility for TF<2.9 by @taehoonlee in #598
• Convert RandomZoom to backend-agnostic and improve affine_transform by @james77777778 in #574
• Add the unstack operation to keras core by @tirthasheshpatel in #597
• Port digit addition rnn example to keras-core by @anas-rz in #614
• Set Jit compile only if Model supports it by @sampathweb in #616
• Port big transfer to keras core by @anas-rz in #615
• Implements H5 legacy saving for Keras Core by @nkovela1 in #605
• Add docstrings in numpy by @guillaumebaquiast in #617

New Contributors

• @james77777778 made their first contribution in #477
• @hazemessamm made their first contribution in #490
• @GallagherCommaJack made their first contribution in #462
• @guillaumebaquiast made their first contribution in #558
• @AmedeoBiolatti made their first contribution in #515
• @pranavvp16 made their first contribution in #582
• @cosmo3769 made their first contribution in #527
• @freedomtan made their first contribution in #600
• @taehoonlee made their first contribution in #598

Full Changelog: v0.1.2...v0.1.3

Keras Core 0.1.2 release

Bug fixes and performance improvements.

What's Changed

• Add adagrad for torch by @haifeng-jin in #548
• Add adamax for torch by @haifeng-jin in #549
• Converted UK Ireland accent recognition to work with Keras Core by @fbadine in #537
• port actor_critic_cartpole to keras core by @anas-rz in #542
• Resolve shape via ops in broadcast_to operation by @sampathweb in #547

New Contributors

• @fbadine made their first contribution in #537

Full Changelog: v0.1.1...v0.1.2

Keras Core 0.1.1 release

Highlights

• TensorFlow dependency removed: it is now possible to use Keras Core with JAX or PyTorch without installing TensorFlow.
• Significant performance improvements for Keras optimizers with PyTorch eager execution. Keras optimizers with the PyTorch backend are now at performance parity with native PyTorch optimizers.
• Added the NumPy backend. Note that it does not support training. You can use it to create and call models / layers, write custom evaluation loops, etc.

What's Changed

• bug fix for torch gpu by @haifeng-jin in #438
• Fix typo in README by @abheesht17 in #441
• Use SeedGenerator backend when creating variable by @sampathweb in #439
• Add names to the input tensor, small typo by @Frightera in #449
• Remove unused variables // adam & adamw by @Frightera in #444
• fixed order of parameters in stateless_apply in JAX distributed example by @martin-gorner in #458
• Added JAX distributed training guide. by @hertschuh in #464
• optimize torch performance by @haifeng-jin in #465
• bug fix for seed generator by @haifeng-jin in #471
• Export DTypePolicy as Policy for backward compat by @abheesht17 in #470
• Make it possible to run Keras Core without TensorFlow. by @fchollet in #475
• Weighted metrics without metrics by @mihirparadkar in #474
• Constant Initializer Error fixed by @kamathis4 in #479
• Complete missing docstrings in keras-core/ops/core.py by @Frightera in #494
• Port zero-dce example to keras-core by @soumik12345 in #486
• Updated math.py & nn.py with docstrings. by @Sheiphan in #484
• Add a requirements file for multi-backend cuda by @mattdangerw in #472
• Stop tensorflow from eating all GPU memory by @mattdangerw in #473
• Adding docstrings to Keras Core ops in numpy.py file by @sukhdeepg in #507
• Fix Torch conv transpose by @chenmoneygithub in #520
• Port mirnet example to keras-core by @soumik12345 in #491
• Testing Timeseries anomaly detection Framework agnostic by @anas-rz in #501
• Added docstring ops/image/resize by @Frightera in #509
• Converted to Keras Core: A Vision Transformer without Attention by @anas-rz in #497
• Optimizer torch optimizer performance by @haifeng-jin in #482
• Clean up variable vs weight in base layer by @mattdangerw in #521
• Convert to Keras_Core: Compact Convolutional Transformer by @anas-rz in #523
• Framework agnostic autoencoder for image deniosing by @anas-rz in #524
• Input layer test by @kamathis4 in #511
• Update symbolic_arguments.py by @sqali in #513
• Orthogonal Inititializer test by @kamathis4 in #526
• Convert to Keras Core: Token Learner by @anas-rz in #528
• Add Adam optimizer parallel implementation for torch backend by @haifeng-jin in #531
• Adding: Numpy Backend by @ariG23498 in #483
• Add AdamW implementation with torch parallel ops by @haifeng-jin in #532
• Fix SavedModel integration and add associated tests by @nkovela1 in #522
• Add RMSprop for Torch by @haifeng-jin in #533
• Add Adadelta for torch by @haifeng-jin in #534

New Contributors

• @abheesht17 made their first contribution in #441
• @Frightera made their first contribution in #449
• @fchollet made their first contribution in #475
• @mihirparadkar made their first contribution in #474
• @kamathis4 made their first contribution in #479
• @soumik12345 made their first contribution in #486
• @Sheiphan made their first contribution in #484
• @sukhdeepg made their first contribution in #507
• @anas-rz made their first contribution in #501
• @sqali made their first contribution in #513

Full Changelog: v0.1.0...v0.1.1

Keras Core 0.1.0 release

Initial preview release. See the announcement post for details on capabilities, credits, etc.