• Added citation information for the accompanying scientific publication in the documentation and the package metadata.

This is a minor release but does contain a range of substantial new features
as well as visual changes, along with some bug fixes. This is accompanied
by internal breaking changes in the R6 classesConverter and
ConvertedModel enabling non-sequential models with multiple input or
output layers. For users, however, nothing changes that is not set by
default as in the previous version or made aware by warnings.

Breaking changes

There are no user-facing changes that are not handled with default values
or noted by throwing warnings.

• When converting a model to a list, two necessary entries are added,
  containing the indices of the layers from the sub-listlayers indicating
  the input (input_nodes) and output (output_nodes) layers of the passed
  model. If one of these values is not set, a warning is raised and it is
  assumed that the model is sequential, i.e., the first layer is the only
  input layer and the last layer is the only output layer.

• Similarly, for each layer in the sub-listlayers the entries
  input_layers andoutput_layers are added containing the indices of the
  preceding and following layers for this layer. If these values are not set, a
  warning is thrown and it is assumed that the model is sequential, i.e., the
  previous entry is the only preceding and the next entry is the only
  succeeding layer. The values0 and-1 indicate the input and output
  layers of the model, respectively.

• The functionsplot andboxplot for the interpretability methods no
  longer return instances ofggplot2 orplotly, but instances of the new
  S4 classesinnsight_ggplot2 orinnsight_plotly. However, these objects can
  be treated like ordinary objects ofggplot2 orplotly to some extent
  and also create the usual visualizations by default (seethis section in the
  in-depth explanation for details). Since the results of models with multiple
  input and output layers are very complex, the suggested packages
  gtable,grid andgridExtra are needed only in these cases.

• Addcli dependency:

  • Errors, warnings, messages, and progress bars have been
    revised and unified, and now use the packagecli.

  • Overwrite the defaultprint() function for the R6 classesConverter and
    InterpretingMethod, which in particular is inherited by all interpretability
    methods.

New features

• TheConverter class now supports more models and layers:

  • Now models created bykeras::keras_model are accepted. In addition, we
    add support for the following layers of thekeras package:
    layer_input,layer_concatenate,
    layer_add,layer_activation*,layer_zero_padding_1d,
    layer_zero_padding_2d,layer_batch_normalization,
    layer_global_average_pooling_1d,layer_global_average_pooling_2d,
    layer_global_max_pooling_1d andlayer_global_max_pooling_2d

  • For models created by thetorch package, we add support for
    nn_batch_norm1d andnn_batch_norm2d

  • For models defined as a named list, we add the entries described in the
    breaking changes and the following layer types
    (see thein-depth explanation for details):

    • type = "BatchNorm" for batch normalization layers
    • type = "GlobalPooling" for all kinds of global pooling layers, i.e.
      maximum or average global pooling
    • type = "Padding" for padding layers
    • type = "Concatenate" for concatenation layer
    • type = "Add" for an adding layer

• Extend the argumentsoutput_idx (in all interpretability methods and the
  corresponding plot and boxplot methods),input_dim,input_names,
  output_dim,output_names (inConverter), which now allow lists of these
  arguments to define them for multiple input or output layers.

• Overwrite the defaultprint() function for the R6 classesConverter and
  InterpretingMethod, which in particular is inherited by all interpretability
  methods.

• Add the S3 functionget_result() for instances of the R6 class
  InterpretingMethod (i.e., also for all inherited methods) that forwards to the
  corresponding class method$get_result().

• In the methodLRP it is now possible to set the rule and the parameter
  individually for each layer type. In addition, for batch normalization layers
  the rule"pass" is added, which skips this type of layer in the backward pass.

• Add the logical argumentwinner_takes_all to the methodsDeepLift and
  LRP to treat maximum pooling layers as an average pooling layer in the
  backward pass.

• Add the logical argumentverbose to all implemented methods to show or disable
  the progress bar.

Documentation and vignettes

• Revise the documentation and use roxygen templates (@template) for almost
  all fields and arguments. These are stored in the folderman-roxygen.

• Revise the introduction vignetteinnsight (vignette("innsight")).

• Add vignette "Example 1: Iris dataset with torch" describing the basic
  usage of the package with tabular data and only numeric features.

• Add vignette "Example 2: Penguin dataset with torch and luz" describing
  a more advanced usage with tabular data containing numerical and categorical
  features.

• Add article "Example 3: ImageNet with keras" describing the usage of the
  package with predefined models inkeras on the ImageNet dataset.

• Add the vignette "In-depth explanation" explaining all methods, arguments
  and possibilities of the package in great detail. This vignette also includes
  the depreciated vignette "Custom Model Definition".

• The vignette "Custom Model Definition" is deprecated.

Minor improvements and bug fixes

• Small speed improvements by using moretorch functions, e.g.,
  torch_clip(x, min = 0) instead of(x > 0) * x

• Some smaller bug fixes

CRAN release

CRAN release