Persistence Initialization: A novel adaptation of the Transformer architecture for Time Series Forecasting

Public implementation of "Persistence Initialization: A novel adaptation of the Transformer architecture for Time Series Forecasting".

Paper

Link

Run code

1. make pull
2. Run ./docker_run python scripts/run.py

Command line options

The default setting can be found in conf.py. Options can be set via the command line, for instance: ./docker_run python scripts/run.py freq=YEARLY model=TRANSFORMER persistence_init=NONE d_model=64 layer=POST_NORM pos_enc=SINUSOIDAL will run a "regular" Transformer model, without any of the proposed changes from our paper.

It is also possible to run a range of experiments by using Hydra's multirun feature: ./docker_run python scripts/run.py -m freq=YEARLY,QUARTERLY,MONTHLY,WEEKLY,DAILY,HOURLY to run a small model for each data frequency.