Keras 2.2.0 (development)

Install the development version with: install_github("rstudio/keras")

  • 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()

Keras 2.1.6 (CRAN)

  • 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

Keras 2.1.5

Keras 2.1.4

Keras 2.1.3

Keras 2.1.2

  • 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)

keras 2.0.9

keras 2.0.8

  • 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.

keras 2.0.6

  • 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

keras 2.0.5

  • Initial CRAN release