problem of the prediction

[liuxingbin]

When I generated the full mask following https://github.com/ChaoningZhang/MobileSAM#demo, I found that using model.eval() differs from the results of not using model.eval(). Interestingly, model.eval() is already settled at https://github.com/ChaoningZhang/MobileSAM/blob/master/mobile_sam/build_sam.py#L88, and I checked model.training is False. What's the problem?