stingrai-widget

A widget to calculate integrated gradients of neural networks for numerical tabular datasets.

Overview

1. Train model, which should be analyzed.
2. Split the data set on the basis of the prediction of the model. into intervals with code_of_project/get_index_in_interval.
3. Calculate explanations for differences between intervals with code_of_project/calculate_ig.
4. Visualize the Integrated gradients with code_of_project/widget.
5. Compare explanations given by StInGRAI and IG_Unsegmented.

Example Output

Visualisation of the explanation given by the Integrated Gradient for DioGenes data set.

Instruction

1. Train your model. You can add the code for this in code_of_project/train_models: The original data can be stored in datasets. Store the data as it is fed into the neural network, the predictions and the labels in models/experiment_name.

2. Split the data set on the basis of the prediction of the model.

   a) Choose your settings in code_of_project/get_index_in_interval/get_index_in_interval_args.py.

   b) Start splitting with get_index_for_StInGRAI.py.

c) With the script construct_ig_unsegmented_paths.py arbitrary intervals can be combined.

3. Calculate explanations for differences between intervals with code_of_project/calculate_ig.

   a) Add the steps your model need to do predictions in code_of_project/prediction_object/PredictionObject.py.

   b) Choose your settings in code_of_project/calculate_ig/calculate_ig_args.py.

   c) Use cd to go to the folder containing the intervals you calculated in step 2.

   d) Type ls -a in the terminal to get all files in the folder.

   e) Adjust the settings in code_of_project/calculate_ig/calculate_ig_args.sh.

   f) Add the files for which the IG should be calculated in the bash scripts in the folder code_of_project/calculate_ig/bash_scripts.

   g) Start the script `calculate_ig_unsegmented.sh or calculate_StInGRAI.sh in the terminal.

4. Visualize the Integrated gradients with code_of_project/widget.

   a) Choose your settings code_of_project/widget/widget_args.py.

   b) Start the visualization with code_of_project/widget/start_widget.py.

5. Comparison of the explanations given by StInGRAI and IG_Unsegmented.

   a) Choose your settings code_of_project/compare_IG_Unsegmented_vs_StInGRAI/get_explanations/get_explanation_args.py

   b) Save explanations by running code_of_project/compare_IG_Unsegmented_vs_StInGRAI/get_explanations/get_explanation.py

   c) Choose your settings code_of_project/compare_IG_Unsegmented_vs_StInGRAI/compare/compare_args.py

   d) Compare explanations regarding spearman_correlation and absolute error with code_of_project/compare_IG_Unsegmented_vs_StInGRAI/compare/compare_IG_Unsegmented_vs_StInGRAI.py

Problems

1. Problem: Can't interact with plots.

   Solution: Using IDE maybe matplotlib interactive mode is blocked. For Pycharm please uncheck Settings | Tools | Python Scientific | Show Plots in Tool window Test matplotlib interactive mode with code_of_project/test/buttons.py.

2. Problem:UserWarning: Matplotlib is currently using agg, which is a non-GUI backend, so cannot show the figure. plt.show()

   Solution: All I had to do was to install tkinter through the Linux bash terminal using the following command: sudo apt-get install python3-tk.

Contact

In case of problems or suggestions please contact jannis.brugger[at]tu-darmstadt[dot]de.