keras3 1.2.0

• Added compatibility with Keras v3.5.0. User facing changes:

  • New functions:
    • op_associative_scan()
    • op_searchsorted()
    • optimizer_lamb()
  • keras$DTypePolicy instances can now be supplied to dtype argument for
    losses, metrics, and layers.
  • Add integration with the Hugging Face Hub. You can now save models to
    Hugging Face Hub directly save_model() and load .keras models directly
    from Hugging Face Hub with load_model().
  • Added compatibility with NumPy 2.0.
  • Improved keras$distribution API support for very large models.
  • Bug fixes and performance improvements.
  • Add data_format argument to layer_zero_padding_1d() layer.
  • Miscellaneous documentation improvements.
  • Bug fixes and performance improvements.

keras3 1.1.0

• Fixed issue where GPUs would not be found when running on Windows under WSL Linux.
  (reported in #1456, fixed in #1459)

• keras_shape objects (as returned by keras3::shape()) gain == and != methods.

• Fixed warning from tfruns::training_run() being unable to log optimizer learning rate.

• Added compatibility with Keras v3.4.1 (no R user facing changes).

• Added compatibility with Keras v3.4.0. User facing changes:

  • New functions:

    • op_argpartition()
    • op_map()
    • op_scan()
    • op_switch()
    • op_dtype()
    • op_lstsq()
    • op_image_hsv_to_rgb()
    • op_image_rgb_to_hsv()

  • Changes:

    • Added support for arbitrary, deeply nested input/output structures in
      Functional models (e.g. lists of lists of lists of inputs or outputs...)
    • Add support for optional Functional inputs.
      • keras_input() gains an optional argument.
      • keras_model_sequential() gains a input_optional argument.
    • Add support for float8 inference for Dense and EinsumDense layers.
    • Enable layer_feature_space() to be used in a {tfdatasets} pipeline even
      when the backend isn't TensorFlow.
    • layer_string_lookup() can now take tf$SparseTensor() as input.
    • layer_string_lookup() returns "int64" dtype by default in more modes now.
    • Layer() instances gain attributes path and quantization_mode.
    • Metric()$variables is now recursive.
    • Add training argument to Model$compute_loss().
    • split_dataset() now supports nested structures in dataset.
    • All applications gain a name argument, accept a custom name.
    • layer_multi_head_attention() gains a seed argument.
    • All losses gain a dtype argument.
    • loss_dice() gains an axis argument.
    • op_ctc_decode(), new default for mask_index = 0
    • All op_image_* functions now use default data_format value
      to config_image_data_format()
    • op_isclose() gains arguments rtol, atol, equal_nan.
    • save_model() gains argument zipped.
    • Bugs fixes and performance improvements.

keras3 1.0.0

• Chains of layer_* calls with |> now instantiate layers in the
  same order as %>% pipe chains: left-hand-side first (#1440).

• iterate(), iter_next() and as_iterator() are now reexported from reticulate.

User facing changes with upstream Keras v3.3.3:

• new functions: op_slogdet(), op_psnr()

• clone_model() gains new args: call_function, recursive
  Updated example usage.

• op_ctc_decode() strategy argument has new default: "greedy".
  Updated docs.

• loss_ctc() default name fixed, changed to "ctc"

User facing changes with upstream Keras v3.3.2:

• new function: op_ctc_decode()

• new function: op_eigh()

• new function: op_select()

• new function: op_vectorize()

• new function: op_image_rgb_to_grayscale()

• new function: loss_tversky()

• new args: layer_resizing(pad_to_aspect_ratio, fill_mode, fill_value)

• new arg: layer_embedding(weights) for providing an initial weights matrix

• new args: op_nan_to_num(nan, posinf, neginf)

• new args: op_image_resize(crop_to_aspect_ratio, pad_to_aspect_ratio, fill_mode, fill_value)

• new args: op_argmax(keepdims) and op_argmin(keepdims)

• new arg: clear_session(free_memory) for clearing without invoking the garbage collector.

• metric_kl_divergence() and loss_kl_divergence() clip inputs
  (y_true and y_pred) to the [0, 1] range.

• new Layer() attributes: metrics, dtype_policy

• Added initial support for float8 training

• layer_conv_*d() layers now support LoRa

• op_digitize() now supports sparse tensors.

• Models and layers now return owned metrics recursively.

• Add pickling support for Keras models. (e.g., via reticulate::py_save_object())
  Note that pickling is not recommended, prefer using Keras saving APIs.

keras 2.15.0

• Default TensorFlow/Keras version installed by install_keras() is now 2.15.
  This is the last Tensorflow version where where Keras 2 is the default.
  To use Keras with Tensorflow v2.16 and up, use the new {keras3} R package.

• Updates to allow both R packages {keras} and {keras3} to be loaded.

• Updates for R-devel (4.4).

keras3 0.2.0

New functions:

• quantize_weights(): quantize model or layer weights in-place. Currently,
  only Dense, EinsumDense, and Embedding layers are supported (which is enough to
  cover the majority of transformers today)

• layer_mel_spectrogram()

• layer_flax_module_wrapper()

• layer_jax_model_wrapper()

• loss_dice()

• random_beta()

• random_binomial()

• config_set_backend(): change the backend after Keras has initialized.

• config_dtype_policy()

• config_set_dtype_policy()

• New Ops

  • op_custom_gradient()
  • op_batch_normalization()
  • op_image_crop()
  • op_divide_no_nan()
  • op_normalize()
  • op_correlate()
  • `

• New family of linear algebra ops

  • op_cholesky()
  • op_det()
  • op_eig()
  • op_inv()
  • op_lu_factor()
  • op_norm()
  • op_erfinv()
  • op_solve_triangular()
  • op_svd()

• audio_dataset_from_directory(), image_dataset_from_directory() and text_dataset_from_directory() gain a verbose argument (default TRUE)

• image_dataset_from_directory() gains pad_to_aspect_ratio argument (default FALSE)

• to_categorical(), op_one_hot(), and fit() can now accept R factors,
  offset them to be 0-based (reported in #1055).

• op_convert_to_numpy() now returns unconverted NumPy arrays.

• op_array() and op_convert_to_tensor() no longer error when casting R
  doubles to integer types.

• export_savedmodel() now works with a Jax backend.

• Metric()$add_variable() method gains arg: aggregration.

• Layer()$add_weight() method gains args: autocast, regularizer, aggregation.

• op_bincount(), op_multi_hot(), op_one_hot(), and layer_category_encoding() now support sparse tensors.

• op_custom_gradient() now supports the PyTorch backend

• layer_lstm() and layer_gru() gain arg use_cudnn, default 'auto'.

• Fixed an issue where application_preprocess_inputs() would error if supplied
  an R array as input.

• Doc improvements.

keras3 0.1.0

• The package has been rebuilt for Keras 3.0. Refer to for an overview
  and https://keras.posit.co for the current up-to-date documentation.

keras 2.13.0

• Default TF version installed by install_keras() is now 2.13.

• Updated layers:

  • layer_batch_normalization() updated signature, with changes to options for distributed training.
  • layer_embedding() gains a sparse argument.

• Fixed deadlock when an R generator was passed to fit(), predict(), and other endpoints.

• When fit(verbose = "auto") is evaluated in the context of a knitr document
  (e.g., quarto or rmarkdown document being rendered), verbose will now
  default to 2, showing one line per epoch.

keras 2.11.1

• Update S3 method formals per new CRAN requirement (r_to_py.keras_layer_wrapper())

keras 2.11.0

• Default TensorFlow version installed by install_keras() is now 2.11.

• All optimizers have been updated for Keras/TensorFlow version 2.11.
  Arguments to all the optimizers have changed. To access the previous
  optimizer implementations, use the constructors available at
  keras$optimizers$legacy. For example, use keras$optimizers$legacy$Adam()
  for the previous implementation of optimizer_adam().

• New optimizer optimizer_frtl().

• updates to layers:

  • layer_attention() gains score_mode and dropout arguments.
  • layer_discretization() gains output_mode and sparse arguments.
  • layer_gaussian_dropout() and layer_gaussian_noise() gain a seed argument.
  • layer_hashing() gains output_mode and sparse arguments.
  • layer_integer_lookup() gains vocabulary_dtype and idf_weights arguments.
  • layer_normalization() gains an invert argument.
  • layer_string_lookup() gains an idf_weights argument.

• Fixed issue where input_shape supplied to custom layers defined with new_layer_class()
  would result in an error (#1338)

• New callback_backup_and_restore(), for resuming an interrupted fit() call.

• The merging family of layers (layer_add, layer_concatenate, etc.) gain the ability
  to accept layers in ..., allowing for easier composition of residual blocks with the pipe %>%.
  e.g. something like this now works:

  block_1_output <- ...
  block_2_output <- block_1_output %>% 
    layer_conv_2d(64, 3, activation = "relu", padding = "same") %>% 
    layer_add(block_1_output)

• model$get_config() method now returns an R object that can be safely serialized
  to rds.

• keras_array() now reflects unconverted Python objects. This enables passing
  objects like pandas.Series() to fit() and evaluate() methods. (#1341)

keras 2.9.0

• New functions for constructing custom keras subclasses:

  • new_model_class()
  • new_layer_class()
  • new_callback_class()
  • new_metric_class()
  • new_loss_class()
  • new_learning_rate_schedule_class().

  Also provided is mark_active(), a decorator for indicating a class method
  should be an active binding (i.e., decorated with Python's @property).
  mark_active() can be used in the new_*_class family of class constructors
  as well as %py_class%.

• r_to_py() method for R6 classes and %py_class% gain support for
  private fields and methods. Any R objects stored in private will only be
  available to methods, and will not be converted to Python.

• New family of functions for controlling optimizer learning rates during training:

  • learning_rate_schedule_cosine_decay()
  • learning_rate_schedule_cosine_decay_restarts()
  • learning_rate_schedule_exponential_decay()
  • learning_rate_schedule_inverse_time_decay()
  • learning_rate_schedule_piecewise_constant_decay()
  • learning_rate_schedule_polynomial_decay()

  Also, a function for constructing custom learning rate schedules:
  new_learning_rate_schedule_class().

• New L2 unit normilization layer: layer_unit_normalization().

• New regularizer_orthogonal, a regularizer that encourages
  orthogonality between the rows (or columns) or a weight matrix.

• New zip_lists() function for transposing lists, optionally matching by name.

• New plot() S3 method for models.

• pydot is now included in the packages installed by install_keras().

• The png package is now listed under Suggests.

• The %<>% assignment pipe from magrittr is exported.

• format() method for keras models (and derivative methods print(), summary(),
  str(), and py_str()):

  • gain a new arg compact. If TRUE (the default) white-space only
    lines are stripped out of model.summary().
  • If any layers are marked non-trainable or frozen, the model summary
    now includes a "Trainable" column, indicating if a layer is frozen.

• freeze_weights() and unfreeze_weights():

  • gain a flexible which argument that can accept layer names (as character strings),
    an integer vector, a boolean vector, or a function that returns a boolean
    when called with a layer. (see updated examples in ?freeze_weights
  • from and to arguments gain the ability to accept negative integers,
    to specify layers counting from the end of the layers list.

• get_weights() gains a trainable argument that can accept TRUE or FALSE,
  allowing for returning only the unfrozen or frozen weights, respectively.

• timeseries_dataset_from_array():

  • R arrays are now cast to the floatx dtype ("float32" by default)
  • start_index and end_index now are 1-based.

• image_dataset_from_directory() gains a crop_to_aspect_ratio argument which
  can be used to prevent distorting images when resizing to a new aspect ratio.

• Layer is deprecated, superseded by new_layer_class().

• load_model_tf() argument custom_objects gains the ability to accept an
  unnamed list (e.g, of objects returned by new_layer_class() or similar).
  Appropriate names for the supplied objects are automatically inferred.

• Fixed an issue where negative values less than -1 supplied to axis
  arguments were selecting the wrong axis.

• get_layer() gains the ability to accept negative values for the index argument.

• Fixed warning from create_layer_wrapper() when the custom layer didn't have
  an overridden initialize or __init__ method.

• Backend functions:

  • k_clip() min_value and max_value gain default values of NULL,
    can be omitted. NULL is taken as -Inf or Inf, respectively.
  • k_squeeze(): axis argument can be omitted, in which case all axes of size 1 are dropped.
  • k_tile(): n argument can now be supplied as a tensor.
  • New function k_unstack().

• KerasTensor objects (e.g, returned by layer_input()) now inherit S3 methods
  for "tensorflow.tensor".

• plot.keras_training_history() no longer issues message
  `geom_smooth()` using formula 'y ~ x' when method = "ggplot2".

• print and related methods for models (format, summary) now accept
  a width argument.

• evaluate(), fit(), and predict() methods for keras Models now default
  to verbose = "auto", with verbosity adjusted appropriately based on calls to
  keras$utils$disable_interactive_logging(), and contexts like
  ParameterServerStrategy.

• install_keras() now accepts version = "release-cpu" as a valid specification.