About the bias term and initialization

[Joechann0831]

Hi, thank you for your sharing. I have two questions about your code in vdsr.py:
1, Why do you set the bias term as False? The official VDSR model contains bias term, is this a trick for a better performance?
2, The initialization you use in this code is
m.weight.data.normal_(0, sqrt(2./n))
Is this the initialization with MSRA method? How about using nn.init?

Thank you very much.

[cjsoft]

when i was implenting vdsr on another dl framework, I found model with bias=false performs very bad on B100 2x. Not sure why.