Fast Neural Style in Caffe

146 views
Skip to first unread message

Paul Delamusica

unread,
May 8, 2017, 1:09:32 PM5/8/17
to Caffe Users
Hi,

  There are a lot of TensorFlow implementations and at least one Torch implementation, but I can't find an implementation in Caffe. Is there any reason for that? Is there anything fundamentally lacking in Caffe for that job?

  Thanks!
Message has been deleted

Paul Delamusica

unread,
Jun 23, 2017, 11:49:37 AM6/23/17
to Caffe Users
Could you elaborate somewhat? 

The paper seems to involve an image encoder for generating the image and a loss "function", which is a VGG network. The loss network is the same as that of Gatys except the batch size > 1. One can get the loss by copying the diff from the loss network and back propagate it through the encoder to train the weights.  

On Friday, June 23, 2017 at 7:19:33 AM UTC-7, Caffe User wrote:
Perhaps because of Fast Neural Style Transfer requires dynamically controlling the behavior of specific layers in the model network which is not a feature available in Caffe. But I'm not sure if Caffe2 has integrated such feature.

Message has been deleted

Paul Delamusica

unread,
Jun 24, 2017, 10:57:58 AM6/24/17
to Caffe Users
That is the only paper as far as I know, plus a supplement. Are there other papers? Like to see more though.

The layers added in VGG are loss layers. Perhaps I don't understand the word "dynamic". Could you explain what layer changes in what way that Caffe could not do? 

Thanks!
Reply all
Reply to author
Forward
Message has been deleted
0 new messages