@beniz This PR is not ready, I have several questions about the integration of multiple class in joligen:
The cls_semantic_nclasses seems to be used only by GAN models, for diffusion, class conditioning uses the f_s_semantic_nclasses parameter. Should we use cls_semantic_nclasses instead? cc @royale
Is it fine to have multiple heads to embed the class, like I did? I'm not satisfied of this solution because it adds complexity to the code, I kept single head for backward compatibility with old models
Check the backward compatibility with train_config, since *_semantic_nclasses is now an array instead of an int
(As @royale suggested, we could sum embeddings of different classes instead of concatening them)
@beniz This PR is not ready, I have several questions about the integration of multiple class in joligen:
cls_semantic_nclassesseems to be used only by GAN models, for diffusion, class conditioning uses thef_s_semantic_nclassesparameter. Should we usecls_semantic_nclassesinstead? cc @royaletrain_config, since*_semantic_nclassesis now an array instead of an int