/disentangle-interventions

Primary LanguagePython

disentangle-mixture-of-interventions

This repository contains the code corresponding to the work done at the research internship at Adobe.

Running Experiments for General-Mixture-Of-Interventions

  1. Checkout the correct branch with git checkout general_mixture
  2. First install all the dependencies using pip install -r requirements.txt
  3. Now to start the experiments and visualizing results:
    1. python jobber.py : For starting the experiment jobs. Please change the experiment_id in the main function to avoid overwriting the result of old experiment. This will run (num_cpu/2) parallel processes.
      • The __main__ function contains all the experiment configuration.
      • Please change following fields to add more experiments:
        • graph_type : ["ER", " SF"] supported (Erdos-Reneyi and Scale-Free)
        • num_nodes : Represents number of node in the graph.
        • node_card : Cardinality of each node
        • num_edge_dist : Only sampling from ["uniform"] distribution over [1,nC2] supported.
        • num_edge_sample : How many different edge number per configuration to be sampled.
        • sparsity : Number of intervention performed on system.
        • num_node_T : Temperature for selection number of node present in each target.
        • pi_dist_type : ["uniform","inverse"] supported for sampleing the mixing coefficient.
        • pi_scale_alpha : alpha for drichlet distribution, (1,inf). Greater scale means sampling from center of simplex.
        • mixture_sample_size : The number of samples taken from mixture distribution.
        • pi_threshold_scale : scale * (1/|S|) for making a mixing coefficient zero if they go below this limit
        • split_threshold : No need to change this. This is for handling numerical error when splitting coefficient.
        • matching_weight : weight x JS_node + (1-weight) x JS_overall for calcualting Jaccard similarity.
        • experiment_id : Unique name for each experiments. The results jsons are stored in the folder with this name.
    2. python plotter.py : For visualizing the results for the experiments. Can be started even when the experiments are going on.
      • experiment_id : The name of the experiments to view the result.

File Organization

  • Data Handling:
    • data_handle.py : Reading bif file, creating the Bayesian Network, creating internvetion graph, sampling from distirbution.
  • VAE Implementation:
    • Dense Latent Space:
      • model.py : Encoder and Decoder model for first order internvetions.
      • run.py : Running the VAE for first order intervetions.
    • Sparse Latent Space:
      • latent_spaces.py : Different Latent Space designs for sparse general order interventions.
      • model_sparse.py : Encoder and Decoder for genral order internvetions.
      • sparse_run.py : Running the VAE for sparse general order internvetions.
  • Non Overlapping Internvetion:
    • non_overlap_intv_solver.py : Implementation of the non-overlapping mixture of internvetions identifiability algorithm.
  • Demo:
    • app.py : Demo-1 shows the accuracy and similarity plot.
    • app2.py : Demo-2 used during the presentation for showing the usecase.
    • app_utils.py : Utilities used the application files.

Running the Code

  1. First install all the dependencies using pip install -r requirements.txt
  2. Running the Existing functionalities
    1. python run.py : For running VAE implementation of first-order-intervention. Please look at the main function for all the hyperparameter configuration.
    2. python sparse_run.py -options : For running VAE implementation of first-order-intervention. See the available options in the sparse_run.py file. Please look at the main function for all the hyperparameter configuration.
    3. python non_overlap_intv_solver.py : For running the identifiability algorithm. Please look at the main function for all the hyperparameter configuration.
    4. python app2.py : To run the app made to showing the use case.