Fine-tuning flickr style: different loss for train and test

16 views

Skip to first unread message

unread,

Nov 7, 2016, 5:44:21 PM11/7/16

to Caffe Users

Hi,

I'm following the "Fine-tuning CaffeNet for Style Recognition on “Flickr Style” Data" tutorial.

After training the network for ~1600 iterations, I ran the "test" command on it.

The printed loss is around ~4.5, but when training the printed loss was ~0.3.

The accuracy seems to be same.

Any reason the loss is different between the two?

Reply all

Reply to author

Forward

0 new messages