ExtFormer

ExtFormer is an open-source library focusing on time series forecasting using transformer variants. The existing transformer models are extended to use the static dataset which is generally useful to make accurate predictions.

The existing transformer models does not take care of sttaic data intrinsicly in the model architenture. The goal here is to extend the models architecture which emable them to use the static data and improve the predictions.

The codebase is motivated from Time Series Analysis (developer of TimesNet).

• [Disclaimer] This is a work in progress.OBBBB

Before creating docker container prepare all the shell scripts in the script/long_forecasting/M5(Divvy)

To do for next version

• add boolean for the static variables
• add static data in the image OR mounted on the dataset pvc
• setup the githu CI/CD to create the docker container and push to dockerhub

create docker container

docker build -t extformer-image ExtFormer

create docker container without cache

docker build --no-cache -t extformer-image ExtFormer

login to docker

docker login

tag the image just created

docker tag extformer-image:latest ramankhurana/extformer-image:latest

push the image to dockerhub

docker push ramankhurana/extformer-image:latest

On K8 cluster

Once git push is successful, github actions will create the docker image and push it to the docker hub.

This new image should be used for setting up the jobs now. New image always have two tags: latest and a version number. Therefore latest will always give the recent one. To get an older version look for the exact version number.

Since docker image is pulled from the dockerhub, there is actually no need to update the main codebase in k8 cluster area. In future the k8 setup and the main code base can be sperated for convinence.

The dataset has be loaded to the /nfs area in the k8 cluster in order the jobs can access those data files.

run the kubernetes job

kubectl apply -f job.yaml

run all the jobs; one job for each shell script

source jobs-all.sh

delete the batch jobs before submitting them again