\>What does your batchsize in train mode?

I have no idea. It wasn't me who trained it (I acquired this model indirectly from someone who asked me to get it running, they didn't train it either). From the fact it doesn't work in eval mode, I'm inferring it was small (possibly 1, but maybe slightly higher).

\>If you have the pytorch model, I would suggest you could try to reestimate the mean and var using your data

I haven't got a lot of data to play with. I've tried running a single patch through it several times in train mode (which I guess achieves this for the mean, but maybe not the variance) and the model does perform better in eval mode, but still not as good as train mode. Maybe I should try with different bach sizes (2,3,4,...), and more than one patch, and see if that works better. However, I'm not sure how that would differ from the previous training (andwould maybe only work if test batch size matched that batch size?). Or, is your suggestion doing something different? I'll have a play with batch size 1 and multiple patches to start with.

Thanks for your input!