How to use MerCBO for a custom dataset?

[fipeop]

Thanks for the great work with MerCBO.

I want to use it for a custom dataset, in which I would like to have:

1. Initialize the model with custom X and Y values and specific discrete space.
2. An external function, so I need to run an iteration of BO at a time, evaluate an external function (a physical experiment) and then update the model.
   Thanks

[aryandeshwal]

Thank you for your interest!

main.py file contains the main interface to the code and it invokes objective functions based on a format similar to the files inMerCBO/MerCBO/experiments directory. I think labs.py is the easiest example to look at.

In order to use a custom dataset, please follow steps given below:

• Create a file in the MerCBO/MerCBO/experiments (let say custom_dataset.py) with a class (let say class Custom_Dataset)

  • The class needs to have an evaluate method that takes an input (torch tensor) and returns the output (torch tensor). You can call your custom external function here.
  • In the __init__ method of the class:
    • self.n_vertices is a list of size equal to the input dimension. For example, since the input dimension in LABS is 40, self.n_vertices is set to np.array([[2] * 40]) (2 just signifies that each variable has 2 choices).
    • self.suggested_init is where the initial X for the model are sampled randomly. It is called in line 61 of main.py to initialize the model. You can define your custom values for x in self.suggested_init.
    • Just copy the rest from labs.py __init__ from line 20 to line 31 in the __init__ of your new class.

• In the main.py file

  • import your new class from MerCBO.experiments.test_functions.custom_dataset import Custom_Dataset.
  • pass an object of your class to kwag_['objective'] similar to line 170 before calling MerCBO method in line 173.

I think everything else should run fine once you make these changes. Please let me know if it doesn't and I will try to provide more pointers.

[fipeop]

Thank you, I will give it a try later this week and let you know.

Is batch optimization possible or this is legacy code from COMBO? i.e. suggesting more than 1 sample by iteration?

[aryandeshwal]

Batch optimization is also possible with parallel thompson sampling. I can provide a batch version if you need. Please let me know. Thanks!