全卷积网络中上采样层的 "learning multiple" 应该是多少?

What should be the "learning multiple" of the up-sampling layers in the Fully Convolutional Network?

我正在尝试训练全卷积网络 (FCN) 以进行密集预测。

paper 的作者提到:

"We initialize the 2× up-sampling to bi-linear interpolation, but allow the parameters to be learned."

当我阅读他们的training prototxt file时,这些层的学习倍数是

我是否应该将此学习倍数更改为 非零值 以让这些层被学习?

谢谢,

引用 shelhamer

​In further experiments​ on PASCAL VOC we found that learning the interpolation parameters made little difference, and fixing these weights gives a slight speed-up since the interpolation filter gradient can be skipped.

因此,如果您想让他们学习,您可以保持 lr_mult 这种方式或更改为非零值。如果需要,您还可以在 solver.prototxt 中设置 lr_policy

详情见this thread in caffe-users group