This PR refactors the models of NeuralCompression. At a high level it does two things:
We remove all duplicate model implementations that are currently implemented in CompressAI.
We add an implementation of HiFiC, which will be later used for the downstream open-sourcing of MS-ILLM.
As we plan to use several variants of the Hyperprior model, we do introduce a new abstract base class for hyperprior models that encapsulates the HiFiC autoencoder. This will make it easier to build a common interface for training loops involving these classes of models.
The test suite has also been updated to test the new HiFiC autoencoder and its functionality, as well as removing old tests that references deleted code.
This PR refactors the models of NeuralCompression. At a high level it does two things:
As we plan to use several variants of the Hyperprior model, we do introduce a new abstract base class for hyperprior models that encapsulates the HiFiC autoencoder. This will make it easier to build a common interface for training loops involving these classes of models.
The test suite has also been updated to test the new HiFiC autoencoder and its functionality, as well as removing old tests that references deleted code.