LatentGranger

requirements

The LatentGranger code is developed with:

• Python 3.9.7
• pytorch 1.9.1
• pytorch-lightning 1.4.7

The used anaconda (tested with v4.10.1) environment with the complete list of libraries is described in environment.yaml.

data

The toy dataset can be generated with

Rscript generate_toy.R 

architectures

beta vae with fully connected

beta vae with convolutional layers

TO FIX

usage

train the autoencoder

## this will train a simple (Granger)-VAE with fully conected layers over the Toy dataset 
## with gamma = 100 nad max lag = 5
python3 main.py -d toy --arch vae -g 100 --maxlag 5  

XAI

The following command will extract the latent representation, average absolute gradients, neural integrated gradients and latent interventions. The output files will be available in the viz/ folder. By default the last checkpoint for the last trained model is used but a specific trained model can be specified with -t (timepoint) and -c (checkpoint name).

python3 explain.py -d toy --extract --nig --grad --save

Acknowledgements

This work was supported by the European Research Council (ERC) Synergy Grant “Understanding and Modelling the Earth System with Machine Learning (USMILE)” under Grant Agreement No 855187.