Introduce AUTO_VI$save_plot() which is the default method of saving a plot by calling save_plot(). This allows user to override the plot saving method if needed.
Introduce a method AUTO_VI$summary() which allows user to get computed statistics provided in AUTO_VI$..str..().
Introduce a method AUTO_VI$plot_pair() which allows user to put the true residual plot and a null plot side-by-side.
Introduce a method AUTO_VI$plot_lineup() which allows user to generate a lineup for manual inspection.
Introduce AUTO_VI$boot_method() which is the default method of generating bootstrapped residuals. This allows user to override the bootstrapping scheme if needed.
Introduce residual_checker() as a new class constructor of AUTO_VI. It has an argument keras_model_name that will be passed to get_keras_model().
Changes
Integrate the AUTO_VI$select_feature() method into AUTO_VI$feature_pca() for clarity. Now the AUTO_VI$feature_pca() method has one more parameter pattern for specifying feature name pattern.
Remove the type parameter and p_value_type parameter from AUTO_VI$p_value() and AUTO_VI$check(), respectively, and unify the p-value formula. Now the p-value is always calculated as mean(c(null_dist, vss) >= vss), where null_dist is a vector of visual signal strength for null residual plots, and vss is the visual signal strength for the true residual plot.
Improve AUTO_VI$feature_pca_plot(). Now the observed point is always displayed on top of other groups.
AUTO_VI$check() and AUTO_VI$lineup_check() now returns self instead of invisible(self) to provide a visible summary of the check result.
get_keras_model() now have an option format to specify the format of the model to download, including “npz”, “SavedModel” and “keras”. The previous version of autovi downloads the pre-trained model in the “.keras”, which could cause backward compatibility issue due to difference in Python or TensorFlow versions. The “SavedModel” format can better handle this aspect but come with a larger file size so it may slow down the model loading process. The “npz” format is the most recommend one, as it will download a Python script to rebuild the model from scratch and load weights from a “.npz” file. This overcomes many of the issues mentioned above.
Bug Fix
Fix a bug in AUTO_VI$vss() that arguments will be passed incorrectly to KERAS_WRAPPER$image_to_array() when a data.frame or a tibble is provided by the user to predict visual signal strength.
Fix a bug in save_plot() where the path argument was not functioning as intended..