Changed Method design

[lorenzo-consoli]

Abandoned the design with BaseMethod and GenerativeFlow in favor of a more flexible, protocol-based, approach.

The core.methods._base was restructured significantly. The main objects there are the following:

• SupportsProtocol, SupportsTraining, SupportsInference: Protocols that define the structural behavior of the protocols.
• ProtocolSpecs and FlowSpecs are used to hold the specification for general and flow-based protocols.
• Abstract protocols, defining the contracts for compute_loss, predict and matching protocols.
• Public base classes, simply wrapping around the abstract ones.
• Mixin and wrapped protocols: These are needed to wrap around the protocols -- for example, as for matched training steps (eg: the ones that need to call match_fn before compute_loss)

The core.methods.library is then structured in two subfolders core.methods.training and core.methods.inference. I implemented the CFMTrainingProtocol and the ODEInference in the respective folders.

Now the Model class will be instantiated using a module, a training_protocol and an inference_protocol. They have to be set at initialization, but can be overridden at call-time.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB