Neural network for Time Series

Realisation of PyTorch Forecasting TemporalFusionTransformer for "FAVORITA" kaggle competition.

Kaggle Team: Yak-Ryl, Leaderboard: 130/606

Deployment

• create .env files:
  • add absolute path to your project

make create_env_files

• create data dirs:

make dirs

• install requierments.txt

pip install -qr requierments.txt

• download data and unzip:

unzip ./src/data/processed/data_main_v3.pickle.zip ./src/data/processed/data_main_v3.pickle
rm ./src/data/processed/data_main_v3.pickle.zip

• download weights if you want to skip training part:

./src/data/check_points/my_favorite_expirement/epoch=29-val_loss=31.6326.ckpt

Optimizing and Training

• optimizing hyperparameters
  • exp_name could be any name
  • accelerator could be gpu or cpu

make tune exp_name=my_favorite_expirement accelerator=cpu   

• training
  • exp_name could be any name
  • use_optim_params
    • if true will use optimized hp from tuning
    • else will use stock parameters from settings.py

make train exp_name=my_favorite_expirement use_optim_params=true       

Submitting

• create submission.csv
  • check_point_name: specify your best check point
  • submission_file_name: name your submission file

make submmit check_point_name=my_favorite_expirement/epoch=29-val_loss=31.6326.ckpt submission_file_name=submission.csv         

PS

If you find this project usefull - leave a star.
Good luck in your submissions.

contacts

Yak
Ryl (Team Leader)