Intervention Effect Functions in Deep Generative Models

This repository contains code to recreate the results from the paper, "Neural Pharmacodynamic State Space Modeling." It also provides the user with scripts to train the deep generative models detailed in the paper (arxiv link forthcoming). We have used PyTorch Lightning to build out our models (see the docs). To recreate the results on the MMRF CoMMpass dataset, first clone the repo and add it in the data folder.

To run this repo,

1. Set up the environment: Run bash requirements.sh. Each time you use this repository, start with conda activate disease_prog.

To run a hyperparameter sweep on a new dataset:

1. Define a data.py file in a new folder: In the data folder, define a new folder for the new dataset. Then, define a load function in a .py file, similar to load_mmrf in data.py under ml_mmrf.
2. Adjust base.py file: In ief_core/models/base.py, add the loading function for the new dataset in the setup() function.
3. Run hyperparameter sweep: To run a hyperparameter sweep, define a config file in ief_core/configs. A test one has been provided. Then, run: python launch_run.py --config [NAME OF CONFIG FILE].

To train a model given a specific set of hyperparameters,

1. Go into correct directory: Go into ief/ief_core,
2. Run command: Run python main_trainer.py --model_name ssm --ttype lin --reg_type l2 --reg_all all --C 0.01 --dim_stochastic 48 --dim_hidden 300 --fold 1 --max_epochs 15000 --dataset mm --inf_noise 0.0 --data_dir /afs/csail.mit.edu/u/z/zeshanmh/research/ief/data/ml_mmrf/ml_mmrf/output/cleaned_mm_fold_2mos_comb3.pkl
3. Specify appropriate hyperparameters: You can specify the hyperparameters of the model as shown above. Please see ief_core/main_trainer.py and ief_core/models/ssm/ssm.py for all options. To specify a path to save a checkpoint (corresponding to model with best validation loss), simply add --ckpt_path [PATH TO OUTPUT FILE] to the command in step 2.

Finally, to recreate the plots in the paper, see examples/model_analyses_final.ipynb.