Repository for the paper : "Time Series Continuous Modeling for Imputation and Forecasting with Implicit Neural Representations" published in TMLR 2024

Folders and files overview

• data folder: Contains a small subpart of the Electricity dataset specifically prepared for classical imputation and forecasting experiments. If you want to apply TimeFlow to other datasets, the input data must have the same structure as the tensors in the Electricity folder (value/gridpoint pairs with the gridpoint between 0 and 1, etc.).
• experiments folder: Includes the necessary files for training the models and performing inference for both imputation and forecasting tasks.
• save_models folder: Serves as a storage location for saving the trained models.
• src folder: Contains the network architecture and the metalearning procedure required for the experiments.

How to run the experiments

First, you should go to the experiments folder, then, it depends if you want to run an imputation experiment or forecast experiment.

Imputation experiment

1. Navigate to the training folder.
2. Open the file "inr_imputation.sh".
3. Choose the desired parameters for training:
   • Select the "draw_ratio" from the options: {0.05, 0.10, 0.20, 0.30, 0.50}.
   • Choose the "version" from the options: {0, 1}.
   • Note: Lower "draw_ratio" values allow for increasing the "sample_ratio_batch" ratio to speed up training.
   • Recommendation: Keep other parameters the same as proposed in the appendix.
4. Run the file using the command: $bash inr_imputation.sh (GPU usage is strongly encouraged for faster training).
5. Once the model is trained, it will be saved in the save_models folder.
6. To perform inference, go to the inference folder and open the "inference_imputation.sh" file.
7. Set the appropriate parameters used for training (dataset name, draw_ratio, epochs, version).
8. Run the file using the command: $bash inference_imputation.sh, and the imputation MAE score will be displayed

Forecast experiment

1. Navigate to the training folder.
2. Open the file "inr_forecast.sh".
3. Choose the desired parameters for training:
   • Select the "horizon" from the options: {96, 192, 336, 720}.
   • Choose the "version" from the options: {0, 1}.
   • Note: Higher "horizon" values allow for decreasing the "horizon_ratio" ratio to speed up training.
   • Recommendation: Keep other parameters the same as proposed in the appendix.
4. Run the file using the command: $bash inr_forecast.sh (GPU usage is strongly encouraged for faster training).
5. Once the model is trained, it will be saved in the save_models folder.
6. To perform inference, navigate to the inference folder and open the "inference_forecast.sh" file.
7. Set the appropriate parameters used for training (dataset name, horizon, epochs, version).
8. Run the file using the command: $bash inference_forecast.sh, and the forecast MAE score will be displayed.