search

jiangsutx / SRN-Deblur

Repository for Scale-recurrent Network for Deep Image Deblurring
http://www.xtao.website/projects/srndeblur/srndeblur_cvpr18.pdf
MIT License
709 stars 184 forks source link

About the kernel size of the deconvolution layer #38

Closed oobbppoo closed 5 years ago

oobbppoo commented 5 years ago

Hi, in your released paper, you said "all kernel sizes are set to 5", but I found that the kernel size of deconvolution layer in decoder part was set to 4, as your code written: deconv2_4 = slim.conv2d_transpose(deconv3_1, 64, [4, 4], stride=2, scope='dec2_4') deconv1_4 = slim.conv2d_transpose(deconv2_1, 32, [4, 4], stride=2, scope='dec1_4') Why did you make this change? Does this make performance better?

jiangsutx commented 5 years ago

We use kernel size 4 for deconvolution layers. I remember we mentioned that in the paper.

Using kernel size 4 and stride 2 can ensure a scale factor 2x. No other special reasons.