Hello everyone! In the paper of Meta-Schedule, it is designed as a modular enough system to incorporate other ways to select the best transformation code during tuning, such as deep reinforcement learning. Thus, how can i use my own cost model or RL search strategy in Meta-Schedule? Is there any tutorail on such development? Thanks a lot.
Hi, I have implemented bayesian optimization as a search strategy in meta schedule. You could have a look at my fork of TVM and compare the changes made. For some more background and findings you could also give this a read. I hope this helps.
1 Like
thank you a lot! Your course project is quite to the point of my confusion and I would benefit from this.