Question about the calculation of prediction probability

[GOZGOXGOYGO]

Question about the calculation of prediction probability

[xiapeng1110]

Hi, Thank you for your great question.

Several clip-based prompt tuning methods, like coop, cocoop, dualcoop, etc., all employed the softmax-like function. To ensure a fair comparison, we also utilized softmax for prediction probability. However, in multi-label tasks, it's possible that using the sigmoid activation function might yield better results. You could explore experimenting with that approach.

[GOZGOXGOYGO]

Thanks for your prompt reply! Actually, to deal with multi-label problem, DualCoOp outputs the predicted probability by calculating softmax value for a positive and negative pair, instead of calculating for all classes. In this way, the output can be regarded as probability for further ASL calculation. However, in your framework, the output probability will differ significantly with different number of classes for the reason of ustilizing softmax function for all classes. So I'm not sure whether it will decrease the performance of your framework. Anyhow, the presented results are quite excellent! :)

[xiapeng1110]

Yes. Your thinking is comprehensive, and indeed, the situation you mentioned could arise. Thank you for your question!