Transfer Learning for Time Series
Summary:
See the recent Medium article for an overview.
Research Papers/Experiments
Transfer learning for time series classification
Reconstruction and Regression Loss for Time-Series Transfer Learning
How flow forecast helps:
Deleting layers when loading weights
Selective freezing of layers via the configuration file
Variable learning rates for different flow forecast layers
Protocol
1. Run initial hyperparameter sweep to get best values for forecast_history, forecast_length, and all other static hyperparameters.
2. Set hyperparameters for transfer learning runs based on this sweep.
3. Pre-train model on large number of counties based on these parameters. Save these weights.
4. Rerun hyperparameter sweep on target county with these saved weights.