search

JimmyChame / SADNet

Pytorch code for "Spatial-Adaptive Network for Single Image Denoising"
128 stars 12 forks source link

initialization method during training #2

Closed mosquitobite closed 4 years ago

mosquitobite commented 4 years ago

Hi, thanks for sharing! I wonder when training conv2d for deformable conv_offset_mask, can the default xaiver uniform initialization method work? In the original paper of DCN, the initilization method is constant zeros. which can prevent loss from exploding. And have you tried effects of different initialization methods for conv_offset_mask learning? Personally, I think initialization method matters since deformable convolution is quite sensitive.

JimmyChame commented 4 years ago

Thanks for your advice. I just use the default initilization method in the DCN, and it can obtain good results. Initialization methods do have an impact on the results, and we will test them more extensively in the future work.