How do you approach when calculating score_map per batch in inference?

[ingbeeedd]

How can you calculate when an image is received in a batch, unlike when inferring from the whole?

[GreatScherzo]

I also have some questions about this. How do you guys set the batch size?
The batch size actually affects the result, as the embedded vectors within the first 3 layers seem to change based on the batch size.