Merging models without access to training data

[MoeinSorkhei]

Hello, thanks for the nice work.
RegMean appears to rely on the training data of individual models for computing inner products. Is there a workaround when training data is not available, such as cases where we have access to pre-trained models but not the pre-training data?