Closed KosumosuL closed 5 years ago
Hi, Yes, you need to compute the loss in each scale. But the ground-truth labels (images) are very easy to obtain, you just need to downsample the images and then compute losses.
thanks :)
Hi, rwenqi! Sorry to be a bother, but i am really interested in your work and want to train your model on my dataset. However i'm confused by the the multi-scale content loss that combined the loss in each scale. Does this means that i need to create a multi-label train set, and each label is the ground truth with different scales? Owing to my poor comprehension of caffe, it's hard for me to solve this problem, can you share your approach with me?
hi, buddy , have you achieved the trained model in your dataset? I am now have some questions for the same purpose, can you leave me a contact or add a wechat for discussing the question, my wechat is mxz1997112.
Actually, the training loss is just the MSE difference between the output and the resized ground truth on each scale, i.e., you do not need to create a multi-label training set, just resize the ground truth when computing the loss.
On Fri, Sep 7, 2018 at 4:20 PM LucasZhao notifications@github.com wrote:
Hi, rwenqi! Sorry to be a bother, but i am really interested in your work and want to train your model on my dataset. However i'm confused by the the multi-scale content loss that combined the loss in each scale. Does this means that i need to create a multi-label train set, and each label is the ground truth with different scales? Owing to my poor comprehension of caffe, it's hard for me to solve this problem, can you share your approach with me?
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/rwenqi/GFN-dehazing/issues/2, or mute the thread https://github.com/notifications/unsubscribe-auth/AOQNvvqOUG_nNvSCgi9waiz7pXByXlncks5uYixQgaJpZM4WeY3_ .
-- Wenqi Ren
School of Computer Science and Technology Tianjin University Homepage: https://sites.google.com/site/renwenqi888/
Hi, rwenqi! Sorry to be a bother, but i am really interested in your work and want to train your model on my dataset. However i'm confused by the the multi-scale content loss that combined the loss in each scale. Does this means that i need to create a multi-label train set, and each label is the ground truth with different scales? Owing to my poor comprehension of caffe, it's hard for me to solve this problem, can you share your approach with me?