You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Adopting your notations in figure 5(d), you initialize the final linear layer W with the text embedding V, and you also keep the visual projection W_v. When you do LP, do you train both V and W_v, or just W? From your code, it seems that you turned off require_grad for visual.proj, so I guess you only trained W.
In section 5.1, you reached the conclusion that "with the proposed language-init method, one can ensure that few-shot performance is always better than zero-shot". This is not precise because if you only train W, the optimization problem becomes convex and the initialization should not matter (of course this is up to the optimizer's inductive bias). So I suspect the reason why you have LP better than zero-shot in figure 6 is due to the inherent regularization from the optimizer (for example, early stopping ~ l2 regularization). For the CLIP paper, they used L-BFGS to solve LP so initialization really didn't matter to them.
The text was updated successfully, but these errors were encountered:
Adopting your notations in figure 5(d), you initialize the final linear layer W with the text embedding V, and you also keep the visual projection W_v. When you do LP, do you train both V and W_v, or just W? From your code, it seems that you turned off require_grad for visual.proj, so I guess you only trained W.
In section 5.1, you reached the conclusion that "with the proposed language-init method, one can ensure that few-shot performance is always better than zero-shot". This is not precise because if you only train W, the optimization problem becomes convex and the initialization should not matter (of course this is up to the optimizer's inductive bias). So I suspect the reason why you have LP better than zero-shot in figure 6 is due to the inherent regularization from the optimizer (for example, early stopping ~ l2 regularization). For the CLIP paper, they used L-BFGS to solve LP so initialization really didn't matter to them.
The text was updated successfully, but these errors were encountered: