Tensor to ndarray: surrogates.utils.sequential_posterior - candidates shape and reshaping

[rjavadi]

I've converted the type of candidates to a NumPy ndarray. However, this change introduced an issue with using flatten(). NumPy's flatten() method doesn't have end_dim argument. Therefore, I need to find an alternative approach for flattening/reshaping the array while preserving the desired structure.

Any clues or suggestions regarding this refactoring is appreciated.

[AVHopp]

Can you point us to the point where exactly this changed should happen?

[AVHopp]

Ahhh I should maybe also read the title of a conversation :D

[Scienfitz]

@rjavadi what are you aiming to achieve with this? the surrogate modeul seems close to the bayesian optimziation and hence it seems should use tensors

[rjavadi]

@Scienfitz In our meetings we discussed that the only surrogate model that requires using tensors is gaussian process, so I changed the base class to use ndarray instead of torch.

I may be wrong :) it's also less work for me if we only need to load torch lazily here 😅

[Scienfitz]

hmm I'm not sure that was correct, I think its more like: anything that is used inside acquisition can keep using torch. While the surrogates are strictly speaking not derived of acquisition, they are only ever needed if acquisition is also used so I guess it would be fine to keep them having torch (lazy)

Re your original question:
Did you try with reshape?
ie a ndarray of shape [2,4,3,4] would normally be flattened to [96] but you could just flatten everything except the last dimensions (according to what end_dim was chosen), for example here end_dim=-2 ignoring the last dimension resulting in shape [24, 4]. I would hope this creates the right order

We need to keep in mind the performance of the code in the gaussian process code, ie it might be that numpy performs copies where torch doesn't or similar so that it would be slower

[rjavadi]

Thanks for reply! I'm going to revert surrogates changes and use torch.

[Scienfitz]

closing as no longer relevant