Hi there,
In this work, I'm sharing my reimplementation for non random initialization for DFL-CNN, which is a part of the paper "Learning a Discriminative Filter Bank within a CNN", just move on to main.py.
I also reimplemented Resnet based DFL-CNN, just check out DFL.py.
Still, there are some issues worth discussing, like "why a special initialization for a certain layer of CNN is able to affect the ultimate categorization result?"
Thanks for @songdejia's work in reimplementing DFL-CNN, I cannot do this without his/her pervious contribution!
If you have any question, plz tell me through issue.
Best,
Ian