about the training code

[yxqlwl]

Hi, sorry for interrupting. Could you please kindly provide the training code as well? Thank you~

[cszn]

See https://github.com/cszn/DnCNN/tree/master/TrainingCodes

[SuhailAliyar]

I am an image processing student. I would like implement this for MRI images. i have these doubts after going through the work

1. About the framework used for training.
2. Much details about the training details.
3. I doubts on overfitting with just 7 layers.
4. Is denoising, deblurring and super resolution integrate together on this work. Or the architecture for 3 operations are different.
5. which is the feature descriptor using here.

It will be very useful for me if you can help me.

Thanks.

[my1347]

@zuie21
1.MatConvNet
2.Refer to DnCNN TrainingCodes
3.Using less layers is not related to overfitting I think.
4.Read the papper
5.feature descriptor? here?

are you a 医生吗

[SuhailAliyar]

We are not getting much info from the read me note. Regards, Suhail Aliyar

…

On 09-Feb-2018, at 2:29 PM, EmperorOfChina ***@***.***> wrote: @zuie21 1.MatConvNet 2.Refer to DnCNN TrainingCodes 3.Using less layers is not related to overfitting I think. 4.Read the papper 5.feature descriptor? here? are you a 医生吗 — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

[XSLXANDY]

Sorry, I'm confused.
Is the training codes the same as the DnCNN? But DnCNN doesn't use HQS for training.
The loss (Equation 10 in paper) is divided into 2 equations (Equation 7 and 8 in paper). Then we need to update z{k+1} by using matconvnet. I have no idea what loss should I choose when I wanna update the net. Besides, I don't know what Mathematical expression the Φ term is.
Hope to get your help. Thanks a lot.

[cszn]

@XSLXANDY IRCNN is a model-based optimization method. It is not an end-to-end training method. IRCNN plugs the CNN denoisers into the HQS inference. So, you only need to train the denoisers.