UDA will cause model selection to use entropy
https://gitlab.kitware.com/darpa_learn/learn/-/blob/master/learn/algorithms/templateDomainNetworkSelection/domainNetworkSelection.py#L181. If uda is being done, target_dataset.labeled_indices() == 0
, meaning it will use entropy, which in my experience often picks the wrong net (resnet34 instead of efficientnet). In this case it shouldn't matter because we're forcing it to be efficientnet, but smth we should consider in the future