1. Paper describes initializing the deconv layer with bilinear filter coefficients and train them. But in the provided train/val.prototxt, we can see lr_mult=0, which means, deconv layer is not trained. Any idea why and how does it affect the accuracy?
? 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.
Keep in mind that there is only one channel per class in this particular architecture, so not that much is there to be learned except perhaps for the spatial extent of the kernel. The results for other data (with more scale variation) or other architectures (with more deconvolution channels and layers) could differ.
2. Previous fcn files used group=21 in the deconv layer. But now, they are removed. Any idea how does it affect the accuracy?
? These are equivalent as long as these parameters are not learned. In the group case, the no. of groups is equal to the no. of channels so that each class is interpolated separately. ?In the no group case, only the "diagonal" of the weight matrix is initialized to the bilinear filter kernels so that each class is likewise interpolated separately with all cross-channel weights set to zero.
?Happy brewing,?
Evan Shelhamer
that is:
conv: N class
deconv:N class
N group