Maskrcnn-benchmark: Dose batch size influence the inference result?

[Tiamo666]

❓ Questions and Help

Hello, Thanks for your great work.
I found that I would get slightly different evaluation result when I use different batch size.
For example, on my own dataset, batch size = 1, I got [email protected] = 0.567, while switch batch size = 4, I got [email protected]=0.566 , batch size = 8 , [email protected] = 0.565.
Does the batch size will influence the inference result?

[fmassa]

This is a bit expected, because with a batch size > 1, all the images will be padded on the sides so that they have the same sizes. This in turn might lead to slightly different results.