Closed HUSTYuChen closed 1 year ago
Hi @HUSTYuChen!
AxClient.complete_trial
. Does this work for you? An alternative would be to implement a custom subclass of Metric
with customization for batch-fetching, but I would just suggest going with the Service API route.The develop API and the service API are actually the same (I'm not sure if I'm understanding this correctly.)
But I can answer this question:
I can combine different models in the service API to achieve the same effect as tutorials of developer API on the website?
Yes, indeed you can! See examples of how to make a generation strategy here: https://ax.dev/tutorials/generation_strategy.html (we also recommend just using our auto-selection logic as illustrated in section 1B in that tutorial; this auto-selection will happen automatically if you use the Service API and don't specify a generation_strategy
argument to AxClient).
So in short, if you use AxClient, you should not need to configure your own generation strategy at all. You can specify multiple objectives in AxClient.create_experiment
, and it will select the qNEHVI acquisition function for you automatically. See example for that here at the top of the multi-objective tutorial: https://ax.dev/tutorials/multiobjective_optimization.html#Using-the-Service-API.
Hi @lena-kashtelyan ! Thank you for answering my questions. I'm actually currently using Service Api to solve the problem. But there are a lot of advanced algorithms on tutorials that are implemented by Develop Api. I originally thought it could only be reproduced by Develop Api. Based on your answer, I now know that I can use Service Api to combine different mods to realize these advanced algorithms. According to the relevant answer you provided me, the current built-in plotting functions are only supported to run in Jupyter Notebook, and it would be nice to be able to run them in Pycharm in the future. Special thanks for taking time out of your busy schedule to answer my questions.
Hi again @HUSTYuChen, happy to help! Let us know if you run into any issues setting up the models you wanted to use, in Service API.
I actually don't know much about PyCharm, we'll look into it. I'll rename this issue accordingly and leave it open but put on the Wishlist, since the PyCharm aspect is the only open part of this I believe.
I've attached my own code to the attachment.