Relationship between classifier_free_dropout (0.2) and classifier_scale (1.0)?

[shileims]

hi author, I noticed that your classifier_free_dropout is 0.2, but what is the reason for setting classifier_scale to be 1.0? Thanks

[ZGCTroy]

classifier_scale is the scale of classifier-free guidance in inference. classifier_free_dropout means to drop condition with the probability 0.1 during training.

you can refer to the paper of Classifier-free Guidance

[shileims]

hi @ZGCTroy , Thanks for your reply. I mean if class_scale is 1.0, does it mean sampling doesn't use the classifer_free_guidence? but training uses the classifier_free_guidance? In terms of the code, uncond_mean is not used if classifier_free_scale is equal to 1.0.

    mean = cond_mean + cfg.sample.classifier_free_scale * (cond_mean - uncond_mean)

Thank you.

[ZGCTroy]

Yes. Specifically, classifier-free guidance is only a technique for condition enhancement in sampling. However, it requires both conditional model and unconditional model during training. Instead of training two models, they propose to train only one conditional model by dropouting condition with fixed probability.

[shileims]

Hi @ZGCTroy , Thank you so much! I just want to clarify you use clasifier-free-guidance during training, not use during sampling. Thank you so much!