More explanation about pilot experiments
zheedong opened this issue · comments
Hi.
I would like to know more details about 'pilot experiments' in SEED v1.
"We conduct two experiments to respectively align discrete representations of VQ-VAE and Beit v2 with OPT2.7B [19] model on CC3M [20] dataset."
In here, what do you mean about 'align'? Do you finetuning OPT 2.7B with VQ-VQE tokens? Or you mean any adaptor between VQ-VAE and OPT? It is great if you tell me more about details.
Thank you.
We freeze OPT 2.7B and train the projection layer, which takes the discrete representations of VQ-VAE and Beit v2 as inputs, using caption loss. Specifically, the discrete representations are fed into the project layer as the inputs of OPT 2.7B.