Perhaps you could try this one? GitHub - AntixK/PyTorch-Model-Compare: Compare neural networks by their feature similarity
It seems to work reasonably well, especially for the same model measured against itself which generates a nice diagonal line. However, for two different models, e.g. ResNet18 vs ResNet101, I haven’t had much luck in uncovering any distinct similarity patterns using CKA. I’m not sure if there is a bug in the CKA code or if the two networks just learn very differently.