search

yannadani / dlgm

Implementation of the paper 'Deep Learning of Graph Matching', CVPR 2018.
MIT License
53 stars 15 forks source link

how to compute the bi-stochastic matrix #2

Open venuszhou opened 5 years ago

venuszhou commented 5 years ago

Hello, In the main.py/class VGG_graph_matching/biStochastic_forward() function:

S = v.view(n,m)
for i in range(N):
        S = torch.mm(S, torch.mm(torch.ones(1,n),S).inverse())
        S = torch.mm(torch.mv(S, torch.ones(m,1)).inverse(), S)
return S

The A.inverse() function need the matrix A to be a square matrix, so this cannot work?

ninodimontalcino commented 5 years ago

Hi,

since we focused on optical flow generation, we had no need for the "biStochastic" function. As you can also notice from the missing batch-wise implementation, this function is not completely implemented which is unfortunately not marked in the code. But we would be happy, if you would complete this missing function and add it to our re-implementation.

Cheers, Nino

venuszhou commented 5 years ago

@ninodimontalcino hey, thanks. I am also curious about that when you experiment on the Sintel dataset, how does the loss change? will it decrease significantly?