The PyTorch models are in this file. Unfortunately, there are a number of duplicated models with/without extra layers and batchnorms, etc. Some actions:
[ ] Get the exact model reference from the paper.
From the paper: The results presented in the main paper are for NNs (both NN1 and NN2) with five layers (four hidden layers) with 128 nodes and rectified linear unit (ReLu) activation functions. Results for different NN archi-tectures are shown in the supplementary information (Figure S2).
[ ] Move the code over to this repo.
[ ] Find the weights for the Python/Fortran models and figure out how to marry them up.
In order to test the Fortran implementation of the NN, it would be great to have the original Python inference code so that we can compare the two for input-output equality. Mentioned in https://github.com/m2lines/convection-parameterization-in-CAM/issues/1
The PyTorch models are in this file. Unfortunately, there are a number of duplicated models with/without extra layers and batchnorms, etc. Some actions: