search

gengshan-y / VCN

Volumetric Correspondence Networks for Optical Flow, NeurIPS 2019.
https://gengshan-y.github.io/VCN/
MIT License
154 stars 25 forks source link

Weight decay during training #11

Closed jytime closed 4 years ago

jytime commented 4 years ago

Hi Gengshan,

In the supplementary material of VCN, you highlight that the weight decay term in PWCNet is removed. May I ask what's the effect of removing it? Is it for faster training or performance boost?

Best Regards, Jianyuan

gengshan-y commented 4 years ago

Hi Jianyuan,

From my view, weight decay is added to regularize the model / reduce overfitting, and VCN is already "regularized" in its structure, and does not need extra regularization. As I can remember in experiments, it performs better without weight decay.

jytime commented 4 years ago

Thanks for your quick reply.