This is a minor release but does contain a range of substantial new features
as well as visual changes, along with some bug fixes. For users, however,
nothing changes that is not set by default as in the previous version or made
aware by warnings. An exception to this are the graphics that are created
using plot(). These now contain a small box with information about the
prediction, the sum of the relevances and the goal of the method.
By default, the plot() method now creates a small box within the plot
with the prediction for the instance and the corresponding class. This info
box also contains the sum of the relevances and, if available, the
decomposition target of the method. Displaying the box can be toggled with
the show_preds argument.
The boxplot() function for the interpretation methods has been
renamed plot_global() due to the inappropriate name (especially for images).
The old method boxplot() can still be used, but throws a warning for image
data and executes the method plot_global() internally.
We have now implemented the following other feature attribution methods:
IntegratedGradient)ExpectedGradient)DeepSHAP)LIME) and Shapley values
(SHAP). Both can be applied to arbitrary models (by providing the
prediction function pref_fun) and wrap the suggested packages lime and
fastshap. However, they can only be applied to models with a single input
and output layer.We have added functions for the initialization of R6 classes. In this way, we don't require prior knowledge of R6 syntax for our package. We implemented the following methods:
convert(...) for Converter$new(...)run_grad(...) for Gradient$new(...)run_smoothgrad(...) for SmoothGrad$new(...)run_intgrad(...) for IntegratedGradient$new(...)run_expgrad(...) for ExpectedGradient$new(...)run_deeplift(...) for DeepLift$new(...)run_deepshap(...) for DeepSHAP$new(...)run_lrp(...) for LRP$new(...)run_cw(...) for ConnectionWeights$new(...)run_lime(...) for LIME$new(...)run_shap(...) for SHAP$new(...)In addition to the output index with output_idx, the new argument
output_label for the output label can now also be specified in order to
calculate or visualize only certain output nodes.
if(require("pkgname")) for suggested packages in examplesThis 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 classes Converter 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.
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-list layers 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-list layers the entries
input_layers and output_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 values 0 and -1 indicate the input and output
layers of the model, respectively.
The functions plot and boxplot for the interpretability methods no
longer return instances of ggplot2 or plotly, but instances of the new
S4 classes innsight_ggplot2 or innsight_plotly. However, these objects can
be treated like ordinary objects of ggplot2 or plotly to some extent
and also create the usual visualizations by default (see this 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 and gridExtra are needed only in these cases.
Add cli dependency:
Errors, warnings, messages, and progress bars have been revised and unified, and now use the package cli.
Overwrite the default print() function for the R6 classes Converter and
InterpretingMethod, which in particular is inherited by all interpretability
methods.
The Converter class now supports more models and layers:
Now models created by keras::keras_model are accepted. In addition, we
add support for the following layers of the keras 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 and layer_global_max_pooling_2d
For models created by the torch package, we add support for
nn_batch_norm1d and nn_batch_norm2d
For models defined as a named list, we add the entries described in the breaking changes and the following layer types (see the in-depth explanation for details):
type = "BatchNorm" for batch normalization layerstype = "GlobalPooling" for all kinds of global pooling layers, i.e.
maximum or average global poolingtype = "Padding" for padding layerstype = "Concatenate" for concatenation layertype = "Add" for an adding layerExtend the arguments output_idx (in all interpretability methods and the
corresponding plot and boxplot methods), input_dim, input_names,
output_dim, output_names (in Converter), which now allow lists of these
arguments to define them for multiple input or output layers.
Overwrite the default print() function for the R6 classes Converter and
InterpretingMethod, which in particular is inherited by all interpretability
methods.
Add the S3 function get_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 method LRP 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 argument winner_takes_all to the methods DeepLift and
LRP to treat maximum pooling layers as an average pooling layer in the
backward pass.
Add the logical argument verbose to all implemented methods to show or disable
the progress bar.
Revise the documentation and use roxygen templates (@template) for almost
all fields and arguments. These are stored in the folder man-roxygen.
Revise the introduction vignette innsight (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 in keras 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.
Small speed improvements by using more torch functions, e.g.,
torch_clip(x, min = 0) instead of (x > 0) * x
Some smaller bug fixes
.Rd files using the current CRAN version of roxygen2.NEWS.md file to track changes to the package.