Statistical Model Criticism of Variational Auto-Encoders (Barkhof & Aziz, 2022)

In this repository I store the code base of my research work conducted at ILLC UvA fall 2021 to spring 2022 in collaboration with Wilker Aziz.

🔗 Link to pre-print paper on Arxiv

A note on the analysis code structure

/analysis

• /bda_models
  • bda_dp_mixture_surprisal_vals.py: analyse surprisal values with DP Mixture (NumPyro model + plot functions)
  • bda_MM_latent_analysis.py: analyse latent samples with DP Mixture (calls Wilker's Pyro Model)
  • bda_pixel_model_mnist.py: conditional beta bernoulli pixel model (NumPyro + plot functions)
  • bda_sequence_length_model_ptb.py: rate poisson model (NumPyro + plot functions)
  • bda_topic_model_ptb.py topic model (uses altered Gensim LDA implementation, optimised with VI)
  • gensim_LDA.py alteration of the class by Gensim
  • /Pyro_BDA: holds code from probabll/bda repository
• /data_space:
  • MNIST_pixels.ipynb: fit BDA & compute surprisals for MNIST
  • PTB_sequence_length_preprocess.ipynb: pre-process for length analyis (outputs ptb_length_analysis_data.pt)
  • PTB_sequence_length.ipynb: fit BDA & compute surprisals for PTB sequence length analysis
  • PTB_topics.ipynb: fit BDA & compute surprisals for PTB lda topic analysis
  • surprisal_DPs.ipynb: fit DPs on surprisal values
  • /output_files: stores intermediate output files
• /latent_space
  • latent_space_analysis.ipynb: perform latent space analysis
  • latent_analysis.py: some functions that help in latent_space_analysis.ipynb
  • image_encoding_stats.csv: stores some basic statistics of the image encodings (KS, MMD, etc.)
  • language_encoding_stats.csv: stores some basic statistics of the language (KS, MMD, etc.)
• final_selection_runs.csv: file that stores the experiments used in the analysis
• global_stats.csv: file that stores aggregated intrinsic evaluation results used in the analysis