TODO

• if mainhydra.py works for both, we can remove mainhydra_turbolence.py (and maybe just rename to main.py)
• Once the dataset is ready, modify the following functions:
  • postprocessing/plots.py: write plot_predictions1D and plot_prediction2D and delete plot_predictions_legacy
  • mainhydra.py: remove lines 50-56 if possible
  • utils/variable_generator.py: remove all file
  • config folder: modify data folder

PACSproject

To use the visualization_turbulent_radiative_layer_2D dataset, you need to install in your environment:

pip install the_well

and then:

the-well-download --base-path path/to/base --dataset turbulent_radiative_layer_2D --split train

where path/to/base should be the path you want it to be installed in (starting from current path)

IDEAS: Cleanliness of the code:

• add abstract classes targetnetwork and hypernetwork? This would be useful to include parameters like replace_weights
• testing function replicates code of training function
• add something ready to use for inference? Future development:
• use a graph structure (ask Riccardo) for connecting different modules of hypernetwork and targetnetwork
• split hypernetwork in backbone and head: -backbone acts as feature extractor: takes hyperparameters and produces a latent space -head maps the latent space to the weights of the targetnetwork: we can design different heads with different initializations straegies to match the needs of the targetnetworks