Implementation of Flamingo, state-of-the-art few-shot visual question answering attention net, in Pytorch. It will include the perceiver resampler (including the scheme where the learned queries contributes keys / values to be attended to, in addition to media embeddings), the specialized masked cross attention blocks, and finally the tanh gating at the ends of the cross attention + corresponding feedforward blocks
Visit Official Website