Unfortunately, despite my preference for jax + flax I had to abandon this project
in favour of (successful) existing and maintained torch implementations (*sigh).
- The implementation as well as experimental setting closely follow the ones described by Bryan Lim, Sercan O. Arik, Nicolas Loeff, Tomas Pfister in their publication Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting
- The codebase is mostly adapted
from google-research/google-research/tft
repository.
If you need a local copy of it, run:
./scripts/clone_original_implementation.sh
srun --partition=cpu-2h --ntasks-per-node=2 --pty bash
apptainer build --fakeroot images/image.sif images/image.def