This paper studies the fast adaptive beamforming for the multiuser multiple-input single-output downlink. Existing deep learning-based approaches assume that training and testing channels follow the same distribution which causes task mismatch, when the testing environment changes. Although meta learning can deal with the task mismatch, it relies on labelled data and incurs high complexity in the pre-training and fine tuning stages. We propose a simple yet effective adaptive framework to solve the mismatch issue, which trains an embedding model as a transferable feature extractor, followed by fitting the support vector regression. Compared to the existing meta learning algorithm, our method does not necessarily need labelled data in the pre-training and does not need fine-tuning of the pre-trained model in the adaptation. The effectiveness of the proposed method is verified through two well-known applications, i.e., the signal to interference plus noise ratio balancing problem and the sum rate maximization problem. Furthermore, we extend our proposed method to online scenarios in non-stationary environments. Simulation results demonstrate the advantages of the proposed algorithm in terms of both performance and complexity. The proposed framework can also be applied to general radio resource management problems.