christopherschmidt89/gpucmiknn

GPU-based information-theoretic conditional independence test for causal discovery

gpucmiknn

A research implementation of a concept for GPU-accelerated information theoretic causal discovery, based upon the CMIknn test for conditional independence ¹. The gpucmiknn implementation can be extended to handle different CMI estimators that build upon the concept of knn searches.

Installation

1. setup conda environment
   • install conda
   • ho to: https://rapids.ai/start.html#rapids-release-selector and configure the desired Rapids environment build command
   • execute the corresponding build command
   • activate the conda environment
2. build cuda code:
   • in parent directory of this repo:
     cd pc_adjacency_search
   • adjust Makefile.config, e.g. (remaining lines can be commented out):
     

   ANACONDA_HOME := $(HOME)/anaconda3/envs/rapids-21.08
   PYTHON_INCLUDE := $(ANACONDA_HOME)/include \
   $(ANACONDA_HOME)/include/python3.7m \
   $(ANACONDA_HOME)/lib/python3.7/site-packages/numpy/core/include/
   PYTHON_LIB := $(ANACONDA_HOME)/lib
   INCLUDE_DIRS := $(PYTHON_INCLUDE) /usr/local/include
   LIBRARY_DIRS := $(PYTHON_LIB) /usr/local/lib /usr/lib
   CUDA_DIR := /usr/local/cuda
   BUILD_DIR := build
   

   • build the code:
     make
   • tested with python 3.8
3. install python dependencies: pip install -r requirements.txt

Execution

In parent directory of this repo:
cd pc_adjacency_search python main.py -i ./data/coolingData.csv --permutations 100 --process_count 1 -a 0.01 --par_strategy 2 -k 7

Parameters

The following options are available:

Parameter	Default	Description
-i		Path to input file in .csv format.
-a	0.05	Sets the significance level used within PC algorithm.
-l	None	Gives the max level for the PC algorithm (level of the pc algorithm is <= max level)
-k	adaptive	k-nearest neighbors during CMI estimation. Adaptive, sets the parameter to 0.2 the sample size.
--permutations	50	Number of Permutations used for the CI Test.
--par_strategy		Flag indicate the parallel hardware used: 1 - CPU-based execution; 2 - GPUKNNCMI-Single; 3 GPUKNNCMI-Parallel
--process_count	2	Number of parallel processes used during adjacency search for CPU-based execution
-b	None	Blocks during block-wise processing of GPUKNNCMI-Parallel. Default setting calculates the blocks on encountering memory pressure due to large numbers of separation set candidates.

Contributor

• Constantin Lange
• Christopher Hagedorn

License

GPL-3

References

Footnotes

1. CMIknn: J. Runge (2018): Conditional Independence Testing Based on a Nearest-Neighbor Estimator of Conditional Mutual Information. In Proceedings of the 21st International Conference on Artificial Intelligence and Statistics. http://proceedings.mlr.press/v84/runge18a.html ↩