High-Performance Open-Source Archive
use_backend("jax") on macOS now defaults to
gpu = FALSE, so jax-metal is no longer
selected automatically.
JAX now registers backend tensor S3 methods for
jax.core.Tracer objects, improving compatibility for traced
computations.
On Linux, tensorflow-cpu is no longer pinned to
2.18.* when resolving Python dependencies.
register_keras_serializable() now updates R layer
wrappers to use the registered class when called.
Numeric ops now include op_layer_normalization(),
op_cbrt(), op_corrcoef(),
op_deg2rad(), op_heaviside(),
op_sparse_sigmoid(), and
activation_sparse_sigmoid().
op_dot_product_attention() gains
attn_logits_soft_cap.
Added signal window operations: op_bartlett(),
op_blackman(), op_hamming(),
op_hanning(), and op_kaiser().
Added loss_categorical_generalized_cross_entropy()
for training with noisy labels.
LoRA-enabled layers (layer_dense(),
layer_embedding(), layer_einsum_dense()) gain
a lora_alpha argument to scale the adaptation delta
independently of the chosen rank.
Added complex-valued helpers: S3 Arg() methods for
tensors, op_angle(), and conversions
op_view_as_real() /
op_view_as_complex().
Added the Muon optimizer via
optimizer_muon().
Added elastic deformation utilities for images:
layer_random_elastic_transform() and the lower-level
op_image_elastic_transform().
Added as.array() support for
PIL.Image.Image objects.
Transposed convolution utilities now follow the latest Keras API:
op_conv_transpose() defaults to strides = 1,
and layer_conv_*_transpose() layers expose
output_padding for precise shape control.
register_keras_serializable() now returns a
registered Python callable, making it easier to use with bare R
functions.
save_model_weights() adds a
max_shard_size argument to split large weight files into
manageable shards.
keras_variable() now accepts a
synchronization argument for distributed
strategies.
layer_layer_normalization() now omits the
rms_scaling argument.
Merging layers now capture ... with tidy dots
(#1525).
JAX _DimExpr shapes now remain symbolic when
combined with R double scalars.
layer_reshape() now accepts -1 as a
sentinel for an automatically calculated axis size.
layer_torch_module_wrapper() gains an
output_shape argument to help Keras infer shapes when
wrapping PyTorch modules.
Layer$add_weight() gains an
overwrite_with_gradient option, and layers now provide a
symbolic_call() method.
Added str() S3 method for Keras
Variables.
JAX arrays now have S3 methods for str(),
as.array(), as.double(),
as.integer(), and as.numeric().
Backend tensors now support base array helpers: t(),
aperm(), and all.equal().
Added pillar::type_sum() for JAX variables and
JaxVariable; str() now covers the new JAX
variable class.
Added training caps via config_max_epochs(),
config_set_max_epochs(),
config_max_steps_per_epoch(), and
config_set_max_steps_per_epoch(). The caps can also be set
via the KERAS_MAX_EPOCHS and
KERAS_MAX_STEPS_PER_EPOCH environment variables. Added
config_is_nnx_enabled() to check whether JAX NNX features
are enabled.
Built-in dataset loaders now accept convert = FALSE
to return NumPy arrays instead of R arrays.
plot(history, theme_bw = TRUE) is now compatible
with ggplot2 3.4.0.
plot(model) now reads DPI from
options(keras.plot.model.dpi = 200) (default is
200).
Reexported reticulate functions: py_help(),
py_to_r(), r_to_py(),
py_require(), and import().
super()$initialize() now works in subclassed Keras
classes, and super() behavior is improved in
subclasses.
use_backend("jax", gpu = TRUE) now declares
dependencies compatible with keras-hub.
Exported named_list().
Switching backends twice in a row now works reliably.
layer_dropout() now preserves
noise_shape as an integer array so length-one shapes are
passed to Keras as iterables.
New op_subset() and x@r[...] methods
enable tensor subsetting using R’s [ semantics and
idioms.
New subset assignment methods implemented for tensors:
op_subset(x, ...) <- value and
x@r[...] <- value
Breaking changes: All operations prefixed with op_
now return 1-based indices by default. The following functions that
return or consume indices have changed: op_argmax(),
op_argmin(), op_top_k(),
op_argpartition(), op_searchsorted(),
op_argsort(), op_digitize(),
op_nonzero(), op_split(),
op_trace(), op_swapaxes(),
op_ctc_decode(), op_ctc_loss(),
op_one_hot(), op_arange()
op_arange() now matches the semantics of
base::seq(). By default it starts, includes the end value,
and automatically infers step direction.
op_one_hot() now infers num_classes if
supplied a factor.
op_hstack() and op_vstack() now accept
arguments passed via ....
application_decode_predictions() now returns a
processed data frame by default or a decoder function if predictions are
missing.
application_preprocess_inputs() returns a
preprocessor function if inputs are missing.
Various new examples added to documentation, including
op_scatter(), op_switch(), and
op_nonzero().
New x@py[...] accessor introduced for Python-style
0-based indexing of tensors.
New Summary group generic method for
keras_shape, enabling usage like
prod(shape(3, 4))
KERAS_HOME is now set to
tools::R_user_dir("keras3", "cache") if
~/.keras does not exist and KERAS_HOME is
unset.
new op_convert_to_array() to convert a tensor to an
R array.
Added compatibility with Keras v3.9.2.
New operations added:
op_rot90()op_rearrange() (Einops-style)op_signbit()op_polar()op_image_perspective_transform()op_image_gaussian_blur()New layers introduced:
layer_rms_normalization()layer_aug_mix()layer_cut_mix()layer_random_invert()layer_random_erasing()layer_random_gaussian_blur()layer_random_perspective()layer_resizing() gains an antialias
argument.
keras_input(),
keras_model_sequential(), and
op_convert_to_tensor() gain a ragged
argument.
layer$pop_layer() gains a rebuild
argument and now returns the removed layer.
New rematerialized_call() method added to
Layer objects.
Documentation improvements and minor fixes.
Fixed an issue where op_shape() would sometimes
return a TensorFlow TensorShape
Fixes for metric_iou(), op_top_k(), and
op_eye() being called with R atomic doubles
Keras now uses reticulate::py_require() to resolve
Python dependencies. Calling install_keras() is no longer
required (but is still supported).
use_backend() gains a gpu argument, to
specify if a GPU-capable set of dependencies should be resolved by
py_require().
The progress bar in fit(), evaluate()
and predict() now defaults to not presenting during
testthat tests.
dotty::. is now reexported.
%*% now dispatches to op_matmul() for
tensorflow tensors, which has relaxed shape constraints compared to
tf$matmul().
Fixed an issue where calling a Metric and
Loss object with unnamed arguments would error.
activation_sparse_plus()activation_sparsemax()activation_threshold()layer_equalization()layer_mix_up()layer_rand_augment()layer_random_color_degeneration()layer_random_color_jitter()layer_random_grayscale()layer_random_hue()layer_random_posterization()layer_random_saturation()layer_random_sharpness()layer_random_shear()op_diagflat()op_sparse_plus()op_sparsemax()op_threshold()op_unravel_index()export_savedmodel()random_shuffle() XLA
compilableactivation_celu()activation_glu()activation_hard_shrink()activation_hard_tanh()activation_log_sigmoid()activation_soft_shrink()activation_squareplus()activation_tanh_shrink()config_disable_flash_attention()config_enable_flash_attention()config_is_flash_attention_enabled()initializer_stft()layer_max_num_bounding_boxes()layer_stft_spectrogram()loss_circle()metric_concordance_correlation()metric_pearson_correlation()op_celu()op_exp2()op_glu()op_hard_shrink()op_hard_tanh()op_ifft2()op_inner()op_soft_shrink()op_squareplus()op_tanh_shrink()callback_backup_and_restore(): Added
double_checkpoint argument to save a fallback
checkpointcallback_tensorboard(): Added support for
profile_batch argumentlayer_group_query_attention(): Added
flash_attention and seed argumentslayer_multi_head_attention(): Added
flash_attention argumentmetric_sparse_top_k_categorical_accuracy(): Added
from_sorted_ids argumentAdded native Flash Attention support for GPU (via cuDNN) and TPU (via Pallas kernel) in JAX backend
Added opt-in native Flash Attention support for GPU in PyTorch backend
Enabled additional kernel fusion via bias_add in TensorFlow backend
Added support for Intel XPU devices in PyTorch backend
install_keras() changes: if a GPU is available, the
default is now to install a CPU build of TensorFlow and a GPU build of
JAX. To use a GPU in the current session, call
use_backend("jax").
get_file() with extract = TRUE
or untar = TRUE, the return value is now the path of the
extracted directory, rather than the path of the archive.Logging is now asynchronous in fit(),
evaluate(), and predict(). This enables 100%
compact stacking of train_step calls on accelerators
(e.g. when running small models on TPU).
on_batch_end, this will disable async logging. You can
re-enable it by adding self$async_safe <- TRUE to your
callbacks. Note that the TensorBoard callback is not considered
async-safe by default. Default callbacks like the progress bar are
async-safe.New bitwise operations:
op_bitwise_and()op_bitwise_invert()op_bitwise_left_shift()op_bitwise_not()op_bitwise_or()op_bitwise_right_shift()op_bitwise_xor()New math operations:
op_logdet()op_trunc()op_histogram()New neural network operation:
op_dot_product_attention()
New image preprocessing layers:
layer_auto_contrast()layer_solarization()New Model functions get_state_tree() and
set_state_tree(), for retrieving all model variables,
including trainable, non-trainable, optimizer variables, and metric
variables.
New layer_pipeline() for composing a sequence of
layers. This class is useful for building a preprocessing pipeline.
Compared to a keras_model_sequential(),
layer_pipeline() has a few key differences:
tf.data, the pipeline will also remain tf.data
compatible, regardless of the backend you use.New argument: export_savedmodel(verbose = )
New argument: op_normalize(epsilon = )
Various documentation improvements and bug fixes.
Added compatibility with Keras v3.5.0. User facing changes:
op_associative_scan()op_searchsorted()optimizer_lamb()keras$DTypePolicy instances can now be supplied to
dtype argument for losses, metrics, and layers.save_model() and load .keras
models directly from Hugging Face Hub with
load_model().keras$distribution API support for very large
models.data_format argument to
layer_zero_padding_1d() layer.Fixed issue where GPUs would not be found when running on Windows under WSL Linux. (reported in #1456, fixed in #1459)
keras_shape objects (as returned by
keras3::shape()) gain == and !=
methods.
Fixed warning from tfruns::training_run() being
unable to log optimizer learning rate.
Added compatibility with Keras v3.4.1 (no R user facing changes).
Added compatibility with Keras v3.4.0. User facing changes:
op_argpartition()op_map()op_scan()op_switch()op_dtype()op_lstsq()op_image_hsv_to_rgb()op_image_rgb_to_hsv()optional Functional inputs.
keras_input() gains an optional
argument.keras_model_sequential() gains a
input_optional argument.float8 inference for Dense
and EinsumDense layers.layer_feature_space() to be used in a
{tfdatasets} pipeline even when the backend isn’t
TensorFlow.layer_string_lookup() can now take
tf$SparseTensor() as input.layer_string_lookup() returns "int64"
dtype by default in more modes now.Layer() instances gain attributes path and
quantization_mode.Metric()$variables is now recursive.training argument to
Model$compute_loss().split_dataset() now supports nested structures in
dataset.name argument, accept a custom
name.layer_multi_head_attention() gains a seed
argument.dtype argument.loss_dice() gains an axis argument.op_ctc_decode(), new default for
mask_index = 0op_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.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.
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.
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.
Update S3 method formals per new CRAN requirement
(r_to_py.keras_layer_wrapper())
Fixed an issue where get_file() would place
incorrectly save files in the current working directory.
(#1365)
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)
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()):
compact. If TRUE (the
default) white-space only lines are stripped out of
model.summary().freeze_weights() and
unfreeze_weights():
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_weightsfrom 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():
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:
min_value and max_value gain
default values of NULL, can be omitted. NULL
is taken as -Inf or Inf, respectively.axis argument can be omitted, in which
case all axes of size 1 are dropped.n argument can now be supplied as a
tensor.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.
Breaking change: The semantics of passing a named list to
keras_model() have changed.
Previously, keras_model() would unname()
supplied inputs and outputs. Then, if a named
list was passed to subsequent
fit()/evaluate()/call()/predict()
invocations, matching of x and y was done to
the model’s input and outpt tensor$name’s. Now, matching is
done to names() of inputs and/or
outputs supplied to keras_model(). Call
unname() on inputs and outputs to
restore the old behavior, e.g.:
keras_model(unname(inputs), unname(outputs))
keras_model() can now accept a named list for
multi-input and/or multi-output models. The named list is converted to a
dict in python. (Requires Tensorflow >= 2.4, Python
>= 3.7).
If inputs is a named list:
call(), fit(), evaluate(),
and predict() methods can also accept a named list for
x, with names matching to the names of inputs
when the model was constructed. Positional matching of x is
still also supported (requires python 3.7+).If outputs is a named list:
fit() and evaluate() methods can
only accept a named list for y, with names
matching to the names of outputs when the model was
constructed.New layer layer_depthwise_conv_1d().
Models gain format() and print() S3
methods for compatibility with the latest reticulate. Both are powered
by model$summary().
summary() method for Models gains arguments
expand_nested and show_trainable, both default
to FALSE.
keras_model_custom() is soft deprecated. Please
define custom models by subclassing keras$Model directly
using %py_class% or R6::R6Class().
Fixed warning issued by
k_random_binomial().
Fixed error raised when k_random_binomial() was
passed a non-floating dtype.
Added k_random_bernouli() as an alias for
k_random_binomial().
image_load() gains a color_mode
argument.
Fixed issue where create_layer_wrapper() would not
include arguments with a NULL default value in the returned
wrapper.
Fixed issue in r_to_py.R6ClassGenerator (and
%py_class%) where single-expression initialize
functions defined without { would error.
Deprecated functions are no longer included in the package documentation index.
Default Tensorflow + Keras version is now 2.7.
New API for constructing RNN (Recurrent Neural Network) layers. This is a flexible interface that complements the existing RNN layers. It is primarily intended for advanced / research applications, e.g, prototyping novel architectures. It allows you to compose a RNN with a custom “cell”, a Keras layer that processes one step of a sequence. New symbols:
layer_rnn(), which can compose with builtin cells:rnn_cell_gru()rnn_cell_lstm()rnn_cell_simple()rnn_cells_stack() To learn more, including how to make
a custom cell layer, see the new vignette: “Working with RNNs”.New dataset functions:
text_dataset_from_directory()timeseries_dataset_from_array()New layers:
layer_additive_attention()layer_conv_lstm_1d()layer_conv_lstm_3d()layer_cudnn_gru() and
layer_cudnn_lstm() are deprecated. layer_gru()
and layer_lstm() will automatically use CuDNN if it is
available.
layer_lstm() and layer_gru(): default
value for recurrent_activation changed from
"hard_sigmoid" to "sigmoid".
layer_gru(): default value reset_after
changed from FALSE to TRUE
New vignette: “Transfer learning and fine-tuning”.
New applications:
application_mobilenet_v3_large(),
application_mobilenet_v3_small()application_resnet101(),
application_resnet152(),
resnet_preprocess_input()application_resnet50_v2(),
application_resnet101_v2(),
application_resnet152_v2() and
resnet_v2_preprocess_input()application_efficientnet_b{0,1,2,3,4,5,6,7}()Many existing application_*()’s gain argument
classifier_activation, with default 'softmax'.
Affected:
application_{xception, inception_resnet_v2, inception_v3, mobilenet, vgg16, vgg19}()
New function %<-active%, a ergonomic wrapper
around makeActiveBinding() for constructing Python
@property decorated methods in
%py_class%.
bidirectional() sequence processing layer wrapper
gains a backwards_layer arguments.
Global pooling layers
layer_global_{max,average}_pooling_{1,2,3}d() gain a
keepdims argument with default value
FALSE.
Signatures for layer functions are in the process of being
simplified. Standard layer arguments are moving to ...
where appropriate (and will need to be provided as named arguments).
Standard layer arguments include: input_shape,
batch_input_shape, batch_size,
dtype, name, trainable,
weights. Layers updated:
layer_global_{max,average}_pooling_{1,2,3}d(),
time_distributed(), bidirectional(),
layer_gru(), layer_lstm(),
layer_simple_rnn()
All the backend function with a shape argument
k_*(shape =) that now accept a a mix of integer tensors and
R numerics in the supplied list.
All layer functions now accept NA as a synonym for
NULL in arguments that specify shape as a vector of
dimension values, e.g., input_shape,
batch_input_shape.
k_random_uniform() now automatically casts
minval and maxval to the output
dtype.
install_keras() gains arg with default
pip_ignore_installed = TRUE.
New family of preprocessing layers. These are the
spiritual successor to the tfdatasets::step_* family of
data transformers (to be deprecated in a future release). Added a new
vignette: “Working with Preprocessing Layers”. New functions:
Image preprocessing:
layer_resizing()layer_rescaling()layer_center_crop()Image augmentation:
layer_random_crop()layer_random_flip()layer_random_translation()layer_random_rotation()layer_random_zoom()layer_random_contrast()layer_random_height()layer_random_width()Categorical features preprocessing:
layer_category_encoding()layer_hashing()layer_integer_lookup()layer_string_lookup()Numerical features preprocessing:
layer_normalization()layer_discretization()These join the previous set of text preprocessing functions, each of which have some minor changes:
layer_text_vectorization() (changed arguments)get_vocabulary()set_vocabulary()adapt()adapt() changes:
layer_text_vectorization() instances were valid.reset_state argument is removed. It only ever accepted
the default value of TRUE.batch_size and steps.%>% (previously returned NULL)get_vocabulary() gains a
include_special_tokens argument.
set_vocabulary():
%>% (previously returned NULL)df_data
oov_df_value) are now subsumed in ....layer_text_vectorization():
output_mode change:
"binary" is renamed to "multi_hot" and
"tf-idf" is renamed to "tf_idf" (backwards
compatibility is preserved).output_mode = "int" would incorrectly return a ragged
tensor output shape.Existing layer instances gain the ability to be added to sequential models via a call. E.g.:
layer <- layer_dense(units = 10)
model <- keras_model_sequential(input_shape = c(1,2,3)) %>%
layer()Functions in the merging layer family gain the ability
to return a layer instance if the first argument inputs is
missing. (affected: layer_concatenate(),
layer_add(), layer_subtract(),
layer_multiply(), layer_average(),
layer_maximum(), layer_minimum() ,
layer_dot())
%py_class% gains the ability to delay initializing
the Python session until first use. It is now safe to implement and
export %py_class% objects in an R package.
Fixed an issue in layer_input() where passing a
tensorflow DType objects to argument dtype
would throw an error.
Fixed an issue in compile() where passing an R
function via an in-line call would result in an error from subsequent
fit() calls. (e.g.,
compile(loss = function(y_true, y_pred) my_loss(y_true, y_pred))
now succeeds)
clone_model() gains a clone_function
argument that allows you to customize each layer as it is
cloned.
Bumped minimum R version to 3.4. Expanded CI to test on all supported R version. Fixed regression that prevented package installation on R <= 3.4
Breaking changes (Tensorflow 2.6): - Note: The following breaking
changes are specific to Tensorflow version 2.6.0. However, the keras R
package maintains compatibility with multiple versions of
Tensorflow/Keras. You can upgrade the R package and still preserve the
previous behavior by installing a specific version of Tensorflow:
keras3::install_keras(tensorflow="2.4.0")
predict_proba() and predict_classes() were
removed.model_to_yaml() and model_from_yaml() were
removed.layer_text_vectorization(pad_to_max_tokens=FALSE)set_vocabulary() arguments df_data and
oov_df_value are removed. They are replaced by the new
argument idf_weights.New Features:
Default Tensorflow/Keras version is now 2.6
Introduced %py_class%, an R-language constructor for
Python classes.
New vignettes:
%py_class%.The keras Python module is exported
Major changes to the underlying handling of custom R6 layer classes.
r_to_py() method is provided for
R6ClassGenerator objects.r_to_py(), without going
through create_layer().KerasLayer is deprecated (new classes should inherit
directly from keras$layers$Layer).KerasWrapper is deprecated (new classes should inherit
directly from keras$layers$Wrapper).create_wrapper() is deprecated (no longer needed, use
create_layer() directly).super in scope that resolves to the Python super class
object.super can be accessed in the 3 common ways:
super()$"__init__"()super(ClassName, self)$"__init__"()super$initialize()super()$`__init__`(...) if
appropriate.supports_masking = TRUE attribute is now supportedcompute_mask() user defined method is now
supportedcall() methods now support a training
argument, as well as any additional arbitrary user-defined
argumentsLayer() custom layer constructor is now lazy about
initializing the Python session and safe to use on the top level of an R
package (#1229).
New function create_layer_wrapper() that can create
a composing R function wrapper around a custom layer class.
Refactored install_keras() (along with
tensorflow::install_tensorflow()). Installation should be
more reliable for more users now. If you encounter installation issues,
please file an issue: https://github.com/rstudio/keras/issues/new
Potentially breaking change: numeric versions supplied without a
patchlevel now automatically pull the latest patch release.
(e.g. install_keras(tensorflow="2.4") will install
tensorflow version “2.4.2”. Previously it would install
“2.4.0”)
pandas is now a default extra packages installed by
install_keras()
pyyaml is no longer a installed by default if the Tensorflow version >= 2.6.
Loss functions:
All the loss functions gain the ability to return a callable (a
keras$losses$Loss instance) if y_true and
y_pred arguments are missing.
New builtin loss functions:
loss_huber()loss_kl_divergence()Metric functions:
All the metric functions gain the ability to return a
keras$metrics$Metric instance if called without
y_true and y_pred
Each metric function is now documented separately, with a common
?Metric topic demonstrating example usage.
New built-in metrics:
metric_true_negatives()metric_true_positives()metric_false_negatives()metric_false_positives()metric_specificity_at_sensitivity()metric_sensitivity_at_specificity()metric_precision()metric_precision_at_recall()metric_sum()metric_recall()metric_recall_at_precision()metric_root_mean_squared_error()metric_sparse_categorical_accuracy()metric_mean_tensor()metric_mean_wrapper()metric_mean_iou()metric_mean_relative_error()metric_logcosh_error()metric_mean()metric_cosine_similarity()metric_categorical_hinge()metric_accuracy()metric_auc()keras_model_sequential() gains the ability to accept
arguments that define the input layer like input_shape and
dtype. See ?keras_model_sequential for details
and examples.
Many layers gained new arguments, coming to parity with the interface available in the latest Python version:
| layer name | new argument |
|---|---|
layer_gru |
time_major |
layer_lstm |
time_major |
layer_max_pooling_1d |
data_format |
layer_conv_lstm_2d |
return_state |
layer_depthwise_conv_2d |
dilation_rate |
layer_conv_3d_transpose |
dilation_rate |
layer_conv_1d |
groups |
layer_conv_2d |
groups |
layer_conv_3d |
groups |
layer_locally_connected_1d |
implementation |
layer_locally_connected_2d |
implementation |
layer_text_vectorization |
vocabulary |
The compile() method for keras models has been
updated:
optimizer is now an optional argument. It defaults to
"rmsprop" for regular keras models. Custom models can
specify their own default optimizer.loss is now an optional argument.run_eagerly,
steps_per_execution.target_tensors and sample_weight_mode must
now be supplied as named arguments.Added activation functions swish and gelu. (#1226)
set_vocabulary() gains a idf_weights
argument.
All optimizer had argument lr renamed to
learning_rate. (backwards compatibility is preserved, an R
warning is now issued).
The glue package was added to Imports
Refactored automated tests to closer match the default installation procedure and compute environment of most user.
Expanded CI test coverage to include R devel, oldrel and 3.6.
set_session and
get_session. (#1046)keras_model eg
name. (#1045)layer_text_vectorization with TensorFlow
>= 2.3 (#1131)text to input_text
in text_one_hot (#1133)text_hashing_trick with
missing values (@topepo #1048)k_logsumexp as it was
removed from Keras (#1137)install_keras now installs a fixed version of h5py,
because newer versions are backward incompatible. (#1142)helper-* file.
(#1173)hdf5_matrix if using TF >= 2.4
(#1175)untar argument to get_file as it
seems to be slightly different from extract (#1179)layer_layer_normalization (#1183)layer_multihead_attention (#1184)image_dataset_from_directory (#1185)ragged argument to layer_input.
(#1193)*_generator deadlocks with recent versions of
TensorFlow (#1197)layer_attention (#1000) by @atroiano.Added layer_dense_features.
Added on_test_*, on_test_batch_*,
on_predict_* and on_predict_* to callback
options.
Search for the right optimizers and initializers on TensorFlow 2.0
Fixed bug in function generators when using models with multiple inputs. (#740)
Added export_savedmodel support for TensorFlow 2.0
(#773)
Fixed bug when using metric_ functions.
(#804)
Allow users to pass additional arguments to
install_keras (#808)
Enabled calling Keras models with R arrays. (#806)
Allow passing data.frames as inputs to Keras models.
(#822)
Fixed bug when passing a fixed validation set to
fit_generator (#837)
Fixed bug when passing a TensorFlow dataset to fit
within a tf$distribute scope. (#856)
install_keras will now install Keras dependencies
(#856). It won’t re-install TensorFlow if it’s already
installed.
Fixed deprecation messages showed with TensorFlow v1.14.
Largely reduced tests verbosity.
Use tf.keras as default implementation
module.
Added AppVeyor to test on Windows.
Added flow_images_from_dataframe function
(#658).
Allow for unknown input_shape in
application_* functions.
Added save_model_tf and load_model_tf
to save/load models in the TensorFlow’s SavedModel format.
Improve handling of timeseries_generator() in calls
to fit_generator()
Add support for input_shape argument to
layer_dropout()
Improve error message for data frames passed to
fit(), etc.
Use 1-based axis indices for k_gather()
Added version parameter to
install_keras() for installing alternate/older
versions
Added activation_exponential() function.
Added threshold parameter to
activation_relu()
Added restore_best_weights parameter to
callback_model_checkpoint()
Added update_freq parameter to
callback_tensorboard()
Added negative_slope and threshold
parameters to layer_activation_relu()
Added output_padding and dilation_rate
parameters to layer_conv_2d_transpose()
Added output_padding argument to
layer_conv_3d_transpose()
Added data_format argument to
layer_separable_conv_1d(),
layer_average_pooling_1d(),
layer_global_max_pooling_1d(), and
layer_global_average_pooling_1d()
Added interpolation argument to
layer_upsampling_1d() and
layer_upsampling_2d()
Added dtype argument to
to_categorical()
Added layer_activation_selu() function.
Added KerasWrapper class and corresponding
create_wrapper function.
Fix issue with serializing models that have constraint arguments
Fix issue with k_tile that needs an integer vector
instead of a list as the n argument.
Fix issue with user-supplied output_shape in
layer_lambda() not being supplied to tensorflow
backends
Filter out metrics that were created for callbacks
(e.g. lr)
Added application_mobilenet_v2() pre-trained
model
Added sample_weight parameter to
flow_images_from_data()
Use native Keras implementation (rather than SciPy) for
image_array_save()
Default layer_flatten() data_format
argument to NULL (which defaults to global Keras
config).
Add baseline argument to
callback_early_stopping() (stop training if a given
baseline isn’t reached).
Add data_format argument to
layer_conv_1d().
Add layer_activation_relu(), making the ReLU
activation easier to configure while retaining easy serialization
capabilities.
Add axis = -1 argument in backend crossentropy
functions specifying the class prediction axis in the input
tensor.
Handle symbolic tensors and TF datasets in calls to
fit(), evaluate(), and
predict()
Add embeddings_data argument to
callback_tensorboard()
Support for defining custom Keras models (i.e. custom
call() logic for forward pass)
Handle named list of model output names in metrics
argument of compile()
New custom_metric() function for defining custom
metrics in R
Provide typed wrapper for categorical custom metrics
Provide access to Python layer within R custom layers
Don’t convert custom layer output shape to tuple when shape is a list or tuple of other shapes
Re-export shape() function from tensorflow
package
Re-export tuple() function from reticulate
package
Indexes for get_layer() are now 1-based (for
consistency w/ freeze_weights())
Accept named list for sample_weight argument to
fit()
Fix issue with single-element vectors passed to text preprocessing functions
Compatibility with TensorFlow v1.7 Keras implementation
Support workers parameter for native Keras
generators (e.g. flow_images_from_directory())
Accept tensor as argument to k_pow()
In callback_reduce_lr_on_plateau(), rename
epsilon argument to min_delta
(backwards-compatible).
Add axis parameter to
k_softmax()
Add send_as_json parameter to
callback_remote_monitor()
Add data_format method to
layer_flatten()
In multi_gpu_model(), add arguments
cpu_merge and cpu_relocation (controlling
whether to force the template model’s weights to be on CPU, and whether
to operate merge operations on CPU or GPU).
Record correct loss name for tfruns when custom functions are
provided for loss
Support for custom constraints from R
Added timeseries_generator() utility
function
New layer layer_depthwise_conv_2d()
Added brightness_range and
validation_split arguments to
[image_data_generator()].
Added support for remove_learning_phase in
export_savedmodel() to avoid removing learning
phase.
Normalize validation data to Keras array in fit()
and fit_generator()
Ensure that custom layers return a tuple from
compute_output_shape()
Added Nasnet and Densenet pre-trained models
New layers layer_activation_softmax() and
layer_separable_conv_1d()
Added amsgrad parameter to
optimizer_adam()
Fix incompatibility with Progbar.update() method in Keras 2.1.4
Models saved via export_savedmodel() that make use
of learning phases can now be exported without having to manually reload
the original model.
Ensure that models saved via export_savedmodel() can
be served from CloudML
Run image data generators with R preprocessing functions on the main thread
Return R list from texts_to_sequences()
Various fixes for use_implementation()
function
Added theme_bw option to plot method for training
history
Support TF Dataset objects as generators for
fit_generator(), etc.
Added use_implementation() and
use_backend() functions as alternative to setting
KERAS_IMPLEMENATION and KERAS_BACKEND
environment variables.
Added R wrappers for Keras backend functions
(e.g. k_variable(), k_dot(), etc.)
Use 1-based axis for normalize function.
Fix issue with printing training history after early stopping.
Experimental support for using the PlaidML backend.
Correct handling for R functions specified in
custom_objects
Added with_custom_object_scope() function.
Automatically provide name to loss function during compile (enables save/load of models with custom loss function)
Provide global keras.fit_verbose option (defaults to
1)
Added multi_gpu_model() function.
Automatically call keras_array() on the results of
generator functions.
Ensure that steps_per_epoch is passed as an
integer
Import evaluate() generic from tensorflow
package
Handle NULL when converting R arrays to Keras
friendly arrays
Added dataset_imbd_word_index() function
Ensure that sample_weight is passed to
fit() as an array.
Accept single function as metrics argument to
compile()
Automatically cast input_shape argument to
applications to integer
Allow Keras models to be composable within model pipelines
Added freeze_weights() and
unfreeze_weights() functions.
Implement export_savedmodel() generic from
TensorFlow package
Convert R arrays to row-major before image preprocessing
Use tensorflow.keras for tensorflow implementation
(TF v1.4)
Added application_inception_resnet_v2() pre-trained
model
Added dataset_fashion_mnist() dataset
Added layer_cudnn_gru() and
layer_cudnn_lstm() (faster recurrent layers backed by CuDNN)
Added layer_minimum() function
Added interpolation parameter to
image_load() function
Add save_text_tokenizer() and
load_text_tokenizer() functions.
Fix for progress bar output in Keras >= 2.0.9
Remove deprecated implementation argument from
recurrent layers
Support for passing generators for validation data in
fit_generator()
Accept single integer arguments for kernel sizes
Add standard layer arguments to layer_flatten() and
layer_separable_conv_2d()
Added image_array_resize() and
image_array_save() for 3D image arrays.
Allow custom layers and lambda layers to accept list parameters.
Expose add_loss() function for custom
layers
Add use_session_with_seed() function that
establishes a random seed for the Keras session. Note that this should
not be used when training time is paramount, as it disables GPU
computation and CPU parallelism by default for more deterministic
computations.
Fix for plotting training history with early stopping callback (thanks to @JamesAllingham).
Return R training history object from
fit_generator()
Rename to_numpy_array() function to
keras_array() reflecting automatic use of Keras default
backend float type and “C” ordering.
Add standard layer arguments (e.g. name,
trainable, etc.) to merge layers
Better support for training models from data tensors in TensorFlow (e.g. Datasets, TFRecords). Add a related example script.
Add clone_model() function, enabling to construct a
new model, given an existing model to use as a template. Works even in a
TensorFlow graph different from that of the original model.
Add target_tensors argument in
compile(), enabling to use custom tensors or placeholders
as model targets.
Add steps_per_epoch argument in fit(),
enabling to train a model from data tensors in a way that is consistent
with training from arrays. Similarly, add steps argument in
predict() and evaluate().
Add layer_subtract() layer function.
Add weighted_metrics argument in compile to specify
metric functions meant to take into account sample_weight
or class_weight.
Enable stateful RNNs with CNTK.
install_keras() function which installs both
TensorFlow and Keras
Use keras package as default implementation rather than tf.contrib.keras
Training metrics plotted in realtime within the RStudio Viewer during fit
serialize_model() and
unserialize_model() functions for saving Keras models as
‘raw’ R objects.
Automatically convert 64-bit R floats to backend default float type
Ensure that arrays passed to generator functions are normalized to C-order
to_numpy_array() utility function for custom
generators (enables custom generators to yield C-ordered arrays of the
correct float type)
Added batch_size and write_grads
arguments to callback_tensorboard()
Added return_state argument to recurrent
layers.
Don’t re-export install_tensorflow() and
tf_config() from tensorflow package.
is_keras_available() function to probe whether the
Keras Python package is available in the current environment.
as.data.frame() S3 method for Keras training
history
Remove names from keras_model() inputs
Return result of evaluate() as named list
Write run metrics and evaluation data to tfruns
Provide hint to use r-tensorflow environment when importing keras
Need mirroring services?
Contact our team at info@vpspulse.com.
Mirror powered by VPSpulse
Infrastructure sponsored by VPSPulse & Secure Payments by ArionPay.