initialization method during training

[mosquitobite]

Hi, thanks for sharing! I wonder when training conv3d for deform_conv_offset, can the default normal 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 deform_conv_offset learning? Personally, I think initialization method matters since deformable convolution is quite sensitive.

[XinyiYing]

Following DCN v1 and v2, we initializes deform_conv_offset by constant zeros. We haven't investigated the effects of different initialization methods for deform_conv_offset learning because constant zeros offset initialization method of D3D can work well in video super-resolution task. We suggest you to make some comparative experiments and publish the results here to let more people know. Thanks for your questions and possible contributions to the work.