NEWS.md

In tensorflow: R Interface to 'TensorFlow'

tensorflow 2.20.0

• Updates for reticulate 1.41 and reticuate::py_require(). New function py_require_tensorflow() which can be called at the start of an R session to enable reticulate to resolve tensorflow. Calling install_tensorflow() in most circumstances is no longer necessary.
• install_tensorflow() installs TensorFlow v2.20 by default.
• Fixed an issue where GPUs would not be found when running on Windows WSL Linux (reported in rstudio/keras3#1456, fixed in #599)
• Fixes for NumPy 2.0 (#601)
• Fixes for R-devel (4.5)

tensorflow 2.16.0

• The package now Suggest 'keras3' instead of 'keras'
• install_tensorflow() installs TensorFlow v2.16 by default.
• If install_tensorflow() detects a GPU on Linux, it will automatically install the cuda package and configure required symlinks for cudnn and ptxax.

tensorflow 2.15.0

• install_tensorflow() installs TensorFlow v2.15 by default
• Added compatibility with the latest release of reticulate (> 1.34).

tensorflow 2.14.0

• install_tensorflow() changes:
• Installs TensorFlow v2.14 by default.
• Now will automatically install the required Nvidia CUDA runtime as a pip package if on Linux and a GPU is detected. You can opt-out by passing install_tensorflow(cuda = FALSE). Aside from the Nvidia driver, no other pre-existing Nvidia CUDA packages are now necessary.
• The configure_cudnn argument is now superseded by the new argument cuda.

• New argument metal, for specifying if the tensorflow-metal pip package should be installed on Arm Macs. Defaults to TRUE on Arm Macs.

• Fixed an issue where as.array() and other methods might fail if the tensor had conversion disabled via r_to_py() or convert = FALSE.

• Fixed an issue where Ops group generic dispatch would error one object was a tensor and the other was a non-tensor Python object (e.g., a numpy array).
• Removed long deprecated symbols: install_tensorflow_extras(), tfe_enable_eager_execution()
• tfestimator generics train() and train_and_evaluate() now warn about their deprecation status when called. The will be removed in a future release.

tensorflow 2.13.0

• install_tensorflow() changes:

  • Installs TensorFlow v2.13 by default now.
  • The envname argument new default is "r-tensorflow". This means that unless the envname argument supplied, install_tensorflow() will now install into the "r-tensorflow" environment, bootstrapping a venv of that name if necessary.
  • gains a new_env argument. If TRUE, any existing environment specified by envname is deleted and created anew. Defaults to TRUE if envname is "r-tensorflow", FALSE otherwise.
  • If running on Linux, now detects if NVIDIA GPUs on Linux are installed, and if so, and installs cuDNN (via pip), configures symlinks for tensorflow to find cuDNN, and emits additional instructions for how to install the necessary CUDA drivers to enable GPU usage. Set new arg configure_cudnn = FALSE to disable.
  • pip_ignore_installed default is now FALSE again.
  • On Arm Macs (M1/M2), the default tensorflow package is once again installed, rather than tensorflow-macos and tensorflow-metal.

• New pillar:type_sum() method for Tensors, giving a more informative printout of Tensors in R tracebacks and tibbles.

tensorflow 2.11.0

• install_tensorflow() now installs TF v2.11 by default.

• as_tensor() now coerces bare R atomic vectors to R arrays before conversion. As a consequence, by default, R atomic double vectors now coerce to 'float64' dtype tensors instead of 'float32'.

• shape() gains the ability to accept vectors of length > 1 in ..., including other tf.TensorShapes. Shapes are automatically flattened.

• Fixed an issue where a ListWrapper object of trackable keras layers (e.g., as part of a keras model) would not convert to an R list.

tensorflow 2.9.0

• Generic method updates:
• New methods: all(), any(), sum(), prod(), min(), max(), mean(), range(), cbind(), rbind(), t(), aperm(), sort(), as.vector(), as.character(), as.raster(), is.infinite(), is.finite(), is.nan()
• ^ will now invoke tf.square() or tf.sqrt() directly when appropriate
• |, &, and ! now cast arguments to 'bool' dtype.
• print() now shows 1d shapes without a trailing commas.

• str() method for tensors now returns only a single compact line; str() on a list of tensors now does something sensible.

• install_tensorflow() now install TensorFlow 2.9 by default.

• install_tensorflow() no longer requires conda on Windows, now works in a regular venv.

• Comparing two partially-defined TensorShape now returns TRUE if each dimension matches. e.g.: shape(NA, 4) == shape(NA, 4) now returns TRUE, previously FALSE.

• Tensors with dtype 'string' now convert to R character vectors by methods as.array() and as.matrix(). (previously they converted to python.builtin.bytes, or an R list of python.builtin.bytes objects)

• as_tensor():

• atomic R integer vectors now convert to 'int32', not 'int64'
• casting between integer and floating dtypes is now done via tf$dtypes$saturate_cast() instead of tf$cast().
• shape argument now accepts a tensor.

• fixed issue where expanding a scalar tensor to an nd-array with shape provided as a tensor would raise an error.

• tf.SparseTensor objects now inherit from "tensorflow.tensor".

tensorflow 2.8.0

• Updated default Tensorflow version installed by install_tensorflow() to 2.8.

• as_tensor() gains a shape argument, can be used to fill or reshape tensors. Scalars can be recycled to a tensor of arbitrary shape, otherwise supplied objects are reshaped using row-major (C-style) semantics.

• install_tensorflow() now provides experimental support for Arm Macs, with the following restrictions:

  • "conda" is the only supported installation method.
  • requests for non-default or older tensorflow versions are not supported.

• install_tensorflow() default conda_python_version changes from 3.7 to NULL.

• tf.TensorShape()'s gain format() and print() S3 methods.

• [ method for slicing tensors now accepts NA as a synonym for a missing or NULL spec. For example x[NA:3] is now valid, equivalent to x[:3] in Python.

tensorflow 2.7.0

• Default Tensorflow version installed by install_tensorflow() updated to 2.7

• Breaking changes:

• shape() now returns a tf.TensorShape() object (Previously an R-list of NULLs or integers).
• [ method for tf.TensorShape() objects also now returns a tf.TensorShape(). Use [[, as.numeric, as.integer, and/or as.list to convert to R objects.
• length() method for tensorflow.tensor now returns NA_integer_ for tensors with not fully defined shapes. (previously a zero length integer vector).

• dim() method for tensorflow.tensor now returns an R integer vector with NA for dimensions that are undefined. (previously an R list with NULL for undefined dimension)

• New S3 generics for tf.TensorShape()'s: c, length, [<-, [[<-, merge, ==, !=, as_tensor(), as.list, as.integer, as.numeric, as.double, py_str (joining previous generics [ and [[). See ?shape for extended examples.

• Ops S3 generics for tensorflow.tensors that take two arguments now automatically cast a supplied non-tensor to the dtype of the supplied tensor that triggered the S3 dispatch. Casting is done via as_tensor(). e.g., this now works: as_tensor(5L) - 2 # now returns tf.Tensor(3, shape=(), dtype=int32) previously it would raise an error: TypeError: `x` and `y` must have the same dtype, got tf.int32 != tf.float32 Generics that now do autocasting: +, -, *, /, %/%, %%, ^, &, |, ==, !=, <, <=, >, >=

• install_tensorflow(): new argument with default pip_ignore_installed = TRUE. This ensures that all Tensorflow dependencies like Numpy are installed by pip rather than conda.

• A message with the Tensorflow version is now shown when the python module is loaded, e.g: "Loaded Tensorflow version 2.6.0"

tensorflow 2.6.0

• Updated default Tensorflow version to 2.6.

• Changed default in tf_function() to autograph=TRUE.

• Added S3 generic as_tensor().

• tfautograph added to Imports

• jsonlite removed from Imports, tfestimators removed from Suggests

• Refactored install_tensorflow().

• Potentially breaking change: numeric versions supplied without a patchlevel now automatically pull the latest patch release. (e.g. install_tensorflow(version="2.4") will install "2.4.2". Previously it would install "2.4.0")

• Removed "Config/reticulate" declaration from DESCRIPTION.

• Setting RETICULATE_AUTOCONFIGURE=FALSE environment variable when using non-default tensorflow installations (e.g., 'tensorflow-cpu') no longer required.

• Users will have to call install_tensorflow() for automatic installation.

• 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.

• Fixed an issue where extra packages with version constraints like install_tensorflow(extra_packages = "Pillow<8.3") were not quoted properly.

• Fixed an issue where valid tensor-like objects supplied to log(x, base), cospi(), tanpi(), and sinpi() would raise an error.

tensorflow 2.5.0

• Updated default Tensorflow version to 2.5.
• Added support for additional arguments in tf_function() (e.g., jit_compile)
• Added support for expm1 S3 generic.
• tfe_enable_eager_execution is deprecated. Eager mode has been the default since TF version 2.0.
• Improved error message in tf_config() on unsuccessful installation.

tensorflow 2.4.0

• Fixed error with use_session_with_seed (#428)
• Added a new set_random_seed function that makes more sense for TensorFlow >= 2.0 (#442)
• Updated the default version of TensorFlow to 2.4 as well as the default Python to 3.7 (#454)

TensorFlow 2.2.0 (CRAN)

• Bugfix with all_dims (#398)

• Indexing for TensorShape & py_to_r conversion (#379, #388)

TensorFlow 2.0.0 (CRAN)

• Upgraded default installed version to 2.0.0.

• Tensorboard log directory path fixes (#360).

• Allow for v1 and v2 compat (#358).

• install_tensorflow now does not installs tfprobability, tfhub and other related packages.

TensorFlow 1.14.1 (CRAN)

• Upgraded default installed version to 1.14.0

• Refactored the install_tensorflow code delegating to reticulate (#333, #341): We completely delegate to installation to reticulate::py_install, the main difference is that now the default environment name to install is r-reticulate and not r-tensorflow.

TensorFlow 1.13.1 (CRAN)

• added option to silence TF CPP info output

• tf_gpu_configured function to check if GPU was correctly

tensorflow documentation built on Aug. 22, 2025, 5:12 p.m.