Improve ExperimentalEmulator

[marcosfelt]

This PR is a complete rewrite of the ExperimentalEmulator. The current implementation has not been giving good results on new training and is very difficult to debug. The debugging challenges are primarily due to the complexity of the training loops and the code organization.

This rewrite depends on pytorch-lightning, which simplifies and organizes training and validation code. It uses the LightningModule to centralize the training loop in the pytorch models and the LightningDataModule to automatically convert datasets + domains into dataloaders for pytorch.

Todo

• Create BayesianRegressor using blitz-bayesian + pytorch-lightning
• Create LightningDataModule
• Set up Emulator with proper saving of model
• Make sure to_dict and from_dict work on Emulator
• Create model registry so custom models can be loaded using from_dict
• Add transforms into LightningDataModule to handle categorical variables
• Calculate evaluation scores on test set (RMSE, MAE)
• Cross validation class for pytorch lightning
• Retrain included models
• Adjust tests as necessary
• Set up to download pretrained models at runtime
• Update docs as necessary

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB