Alternatives for a simple log10 transform to DM density fields?

[Yash-10]

This line:
https://github.com/Yash-10/ViT-LSS/blob/c05dce79c31556a63fa12b55a2fe0f59fb797d31/scripts/utils.py#L16
may not work well with large grid sizes (e.g., 256^3 or higher) since those fields tend to have zero pixels in them during mass assignment. Eventually, it can cause NaN pixels during preprocessing and potentially bewilder the model. This is indeed what seems to happen currently when I transfer-learn the model on 256^3 fields in which case I get some weird NaN outputs.

For smaller grid sizes, this issue is, perhaps, a non-issue.