This analysis uses a GPU enabled and cloud scalable pytorch container base image. The pytorch image is modified by installing gcc graphviz graphviz-dev nano vim, and the python packages CausalNex as well as its associated dependencies and pygraphviz which is used to visualize the networks.
-
A description of the most current analysis can be found written as a Jupyter Notebook in GoogleColab here.
-
NOTE: the Colab Notebook will not have the CPU power to run the structure learning algorithm, it is very computationally intensive (~4 hrs on a server with >250 cores).
-
Optionally this could use CyVerse to host a notebook with the analysis data??
Run the causalnex container image, in the detached state -d, by mounting your current working directory (assumes the docker user is running the process within this repo folder) as the folder /work inside the spinning container. Executing the script for structure learning
docker run -d --rm --gpus all -v $(pwd):/work -w /work rbartelme/pytorch-causalnex:0.0.2 python /work/scripts/learn-structure.py
- The experiment
/outputs/Apr06/had many nodes, with weak connections, no pickling of structure model - On the other hand,in
/outputs/May06/switching from row-wise to column wise removal ofNaNremoves most of the dataset - Rerunning with previous
NaNremoval behavior on May 07 2021