question about test about caffe-vdsr HOT 7 CLOSED

Hi, have you installed matconvnet?

from caffe-vdsr.

Hello. First of all, thanks for your wonderful project.
I have a question about the project. Have you tried to train VDSR model using matconvnet??
I have struggled with training it in matconvnet.
There is no gradient clipping method in matconvnet so I clipped weight gradients using L2 norm.
However, after training, PSNR result could not reach to even 32dB.
this is my implementation of gradient clipping in matconvnet DAG version

T = net.params(p).der;
clip = 1 / net.params(p).learningRate;
norm = sqrt(sum(T(:).^2));
if norm>clip
net.params(p).der = T*(clip/norm);
end

I tried various clip values but all trained result were similar.
I wish to know what different or problem are between my training method and Caffe.
If you have any tips or comments, I would be glad. thank you

from caffe-vdsr.

@cheongjunyoung Hi, I didn't trained VDSR using matconvnet.
But I think you can simply implement gradient clip by a clamp-like function, like

grad(>threshold) = threshold
grad(<-threshold) = -threshold

You don't need to do gradient clipping like caffe. The purpose of gradient clipping is to overcome the problem of gradient explosion. Just clip big gradient to a certain value and you can handle this problem.

from caffe-vdsr.

I have installed matconvnet. It seems that I need to run setup before I run a new project.
Thank you for your answering,I have a question about the training of multiple scale.How to train a model with {2,3}or{3,4}?Does it mean the input images are interpolated to different scale???

from caffe-vdsr.

@luciaL Yes, but I didn't achieve better performance than single scale training. You can have a try.

from caffe-vdsr.

Okay，thank you.The input images are divided into the same size patches.Then the output images are also the same size patches.How to get the full image with the same scale factor??

from caffe-vdsr.

It's the same with single scale training. I didn't achieve better performance in my experiments. Maybe you can ask the author of VDSR paper for help.
Thanks.

from caffe-vdsr.