A monorepo for causality-related experiments.
Data is stored "centrally" in the data/ folder.
Data is not committed to the repo (some of it is tens of MB and no-one likes a long repo clone), but rather provide Make targets to download it.
Experiments map roughly to "playing with a relevant library" and have their own directories inside the experiments/ directory.
Notebooks should be written such that the notebook server is run from the relevant subdirectory (for instance, the dowhy notebook server should be run from causality-experiments/experiments/dowhy).
Each experiment subdirectory should have its own environment file, and use it's own (python) environment, to avoid library version conflicts.
First, clone the repo.
Then, run make dirs to make the data directory (which is .gitignored).
To fetch the data, run:
make news
make churn
make housing
make bikes
make london-bikes
make student(Unfortunately, churn, housing and london-bikes will just give you instructions on where to find the data).
- causal discovery - playing with the CausalDiscoveryToolbox
- causalnex - trying the new CausalNex library
- dowhy - trying to use DoWhy on a couple of datasets
- invariant causal prediction - toying with the R libraries accompanying the papers Causal inference using invariant prediction: identification and confidence intervals and Invariant Causal Prediction for Nonlinear Models
- invariant risk minimization - experiments around Invariant Risk Minimization
"Doing" an experiment in this case means making the code Chris wrote runnable (so, providing reqs file, adding make target for data etc.)
- invariant risk minimization experiments
- dowhy experiments
- causal discovery toolbox experiments
- invariant causal prediciton experiments
- move deconfounder repo to an experiment here
This is a repo for playing with things, understanding libraries and experimenting. The code is currently unlicensed, and we accept no liability related to its use.