search

csyanbin / TPN-pytorch

Pytorch Code for ICLR19 paper: Learning to Propagate Labels: Transductive Propagation Network for Few-shot Learning.
175 stars 29 forks source link

A question about the adjacency matrix #2

Closed HX-idiot closed 4 years ago

HX-idiot commented 4 years ago

Hello, it is really a good work! I am confused of the value of diagonal when constructing the adjacency matrix, in your code, you didn't mask the diagonal, which will make the value of diagonal near to 1. does that limit the result of transductive (S is near to Identity matrix under this condition)?

csyanbin commented 4 years ago

Hi, thanks for your interest in our paper. This is the same issue as here: https://github.com/csyanbin/TPN/issues/2#issuecomment-463913199

Currently, we haven't experimented with Wii=0. In LLGC (zhou et al,2004), they set Wii=0 to avoid self-reinforcement. But in our paper, we think it not necessary to prevent this self-reinforcement.

I will try the setting with Wii=0 in the future.