which way is better?why?

[xuanxuanxuan123456]

I'm confused about the implementation method.
In the paper, the Lie algebra part is placed after the output of the final stage of ResNet, and contrastive learning is conducted in the Lie algebra space. However, in the code, the code for handling the Lie algebra space is placed in the dataset.py file, which seems to be a kind of data preprocessing. The model is responsible for extracting "predicted features" from the original image, and then uses L1 loss (add_loss) to minimize the difference between the predicted features and the target features.
Could you tell me which way is better and the reason about it?