This is the source code of our dropout-bayes-based classifier framework.

• run traintest.train_model() to train a neural network classifier (NN).
• run traintest.pipeline() to attack the NN, save the NN predictions, estimate likelihood of NN predictions and perform cumsum & bayesian inference.
• run analysis.plot_heatmap() to plot the accuracy heatmaps.

• run traintest.train_model() to train a neural network classifier (NN).
• run traintest.pipeline() to attack the NN, save the NN predictions, estimate likelihood of NN predictions and perform cumsum & bayesian inference.
• run analysis.plot_epsilon_accrt() to plot the relationship between ϵ and Acc/RT.

• install textattack

  • in my case, I install version==0.3.0 from source codes to avoid some unnecessary dependence
  • further more, in requirements.txt, scipy==1.4.1 is unnecessary and could raise error while during installation, you can replace it with scipy>=1.4.1
  • again, numpy<1.19.0 is unnecessary, numpy==1.21.4 works fine in my case

• run nlp.attack_model_textattack()

• In case of a download timeout of universal-sentence-encoder_4, here is a work-around:

  • open /home/yourusername/.local/lib/python3.9/site-packages/textattack/constraints/semantics/sentence_encoders/universal_sentence_encoder/universal_sentence_encoder.py
  • replace tfhub_url = "https://tfhub.dev/google/universal-sentence-encoder/4"
  • with tfhub_url = './data/universal-sentence-encoder_4/'
  • download universal-sentence-encoder_4.tar.gz from https://tfhub.dev/google/universal-sentence-encoder/4
  • extract the downloaded file in ./data/, so there will be a DDDM/data/universal-sentence-encoder_4/saved_model.pb after the extraction
  • now try running nlp.attack_model_textattack()

• run traintest.pipeline() and analysis.plot_heatmap() to get heatmaps

• run attack_model_artasr() after training the models
• run traintest.pipeline() and analysis.plot_heatmap() to get heatmaps