Fine tuning results on caffe flickr style

[Arka Sadhu]

Hi

I tried out the following tutorial example from caffe : http://caffe.berkeleyvision.org/gathered/examples/finetune_flickr_style.html. I replicated all the steps from it, but the final accuracy differs quite a lot. The tutorial states that the final model achieves an accuracy of 39%, while mine reached an accuracy of 25% at the end of 100k iterations. Has someone else tried the same tutorial and got similar results? I am attaching a screenshot of the same. 

Also on another note : shouldn't the TestNet Output #1: loss (3.8701) be much lower than the one obtained? 

Thanks