I wonder which dataset is the pretrained model in pytorch? ILSVRC-2012? ILSVRC-2014? or others?
I wonder why the top-5 and top-1 error is significantly higher than the same model in caffe?
I wonder why the top-5 and top-1 error is significantly higher than the same model in caffe?
Which model do you have in mind?
For example, in https://github.com/BVLC/caffe/wiki/Model-Zoo#resnets-deep-residual-networks-from-msra-at-imagenet-and-coco-2015, the top-5 of vgg-16 is 7.4%, but 9.62% and 8.5%(with bn) in pytorch.
I’m not sure, but I assume the site reports the multi-scale evaluation instead of the single-scale one.
You can see the different error rates in Table 3 and 4 in their paper.
Table 3 comes pretty close to the pre-trained models.
Thank you for your reply!