关于代码promt generation的问题 about fptrans

[DreamerrW]

期待您的回答！

[Jarvis73]

@DreamerrW
G is the size of a learnable token. The learnable token tensor (self.prompt_tokens) is designed to have the total shape of [bank_size, G, embed_dim], where bank_size denotes the token pool size, and each token in the pool has a size of [G, embed_dim]. In other words, one can think of a "token" as a group of vectors where G is the number of vectors.

For each foreground/background mean feature with size [embed_dim], we broadcast it to [G, embed_dim], and then we can add it to a randomly selected learnable token.

[DreamerrW]

Thanks!