I think it's kind of disingenuous maybe to claim such improvements in training efficiency when they rely on: - Existing models for data pseudo-labelling - Ima

I think it's kind of disingenuous maybe to claim such improvements in training efficiency when they rely on: - Existing models for data pseudo-labelling - ImageNet pretraining - A frozen text encoder - A frozen image encoder