A Good Prompt Is Worth Millions of Parameters: Low-resource Prompt-based Learning for Vision-Language Models