no gradient for losses: batch_sep, batch_con, and batch_spa

[cwangrun]

Hello, thanks for the great work! I find there is no gradient for the proposed losses when using ViT as the vision backbone, I guess there is some issues in explainer.interpret().