Compatibility between RGPE and custom botorch model in AX

[Kh-im]

Hello,

Is the RGPE strategy (meta_learning_with_rgpe) compatible with custom_botorch_model_in_ax ?

I tried to add the :
target_model = get_fitted_model(train_X, train_Y, train_Yvar)
model_list = base_model_list + [target_model]
rank_weights =compute_rank_weights(train_X,train_Y,train_Yvar,base_model_list,target_model,NUM_POSTERIOR_SAMPLES,)

in the init phase of the Custom class, it doesn't seem to be sufficient.

Thanks for your help

[Balandat]

Hmm good question; so the RGPE model should in principle be compatible with the Modular BoTorch model interface, but it would require registering some custom fitting procedures. The custom_botorch_model_in_ax content is somewhat outdated (we removed it in more recent versions).

Could you please take a look at the "Using custom models and acquisition functions
" section of https://github.com/facebook/Ax/blob/main/tutorials/modular_botorch/modular_botorch.ipynb ? This should get you started on this.