Adding a new layer type

New layers should have a Python component and a Fortran component.

Fortran component

The Fortran implementation of the layer should go in ennuf/_internal/fortran/neural_net_mod.f90. Typically, this means adding a new subroutine to the module.

Python component

The layer class

This is our internal representation of the layer that acts as a middle step between the Python implementation (in Keras, PyTorch, etc.) and Fortran.

In ennuf/_internal/ml_model/layers you should add yourlayer.py, containing a class extending the layer abstract base class BaseLayer, so YourLayer(BaseLayer).

layer classes are NOT implementations of those layers. For example, a Dense layer class should not implement any matrix multiplication logic for doing forward passes through a neural network, applying activations, backward passes during training, etc.

It simply holds the minimum amount of information to recreate that layer in our Fortran implementation (number of neurons, activation type, etc); and has methods for automatically generating that Fortran.

All abstract methods must be implemented.

Translation

In addition to the above, we need to implement a way to translate between supported Python machine learning libraries and our internal representation of the layer.

When adding a new layer, support needs to be added in ennuf/_internal/translation. This should detect if a matching layer type from the external library is used, and extract all relevant information to create an instance of the layer class above.