search

Thinklab-SJTU / ThinkMatch

A research protocol for deep graph matching.
https://thinkmatch.readthedocs.io/
Other
833 stars 118 forks source link

Hi, where can I find the code of Algorithm 1 in your paper? #15

Closed Ha0Tang closed 4 years ago

rogerwwww commented 4 years ago

Thank you for your interest.

https://github.com/Thinklab-SJTU/PCA-GM/blob/master/PCA/model.py#L72 from L72 to L81.

Ha0Tang commented 4 years ago

Hi, thanks. What's the tensor shape of emb1, emb2, emb1_new, and emb2_new?

Ha0Tang commented 4 years ago

My task is, input two feature maps F1, F2 (both are N X C X H X W) and aim to output two updated feature maps F1', F2' (both are still N X C X H X W). I would like to use Algorithm 1 proposed in your paper to do this task. Do you have any suggestions? Thanks.

rogerwwww commented 4 years ago

The shapes of emb1, emb2, emb1_new, and emb2_new should be (N x H x C) where N is the number of batches, H is the number of nodes in one graph and C is the size of the feature channel.

For your task, you may transpose and reshape your F1, F2 into (N x HW x C) (if I don't get these notations wrong), and it should work with my implementation.

Ha0Tang commented 4 years ago

Thanks, and what ns_scr and ns_tgt in Eq. s = self.voting_layer(s, ns_src, ns_tgt) stand for?

rogerwwww commented 4 years ago

ns_scr and ns_tgt stands for the number of nodes in two graphs, as we allow different numbers of nodes in single batch.

Ha0Tang commented 4 years ago

We have HW nodes in each graph, so ns_scr and ns_tgt should be set HW? When I am carrying out s = self.voting_layer(s, H*W, H*W), I got the following error: for b, n in enumerate(nrow_gt): TypeError: 'int' object is not iterable

rogerwwww commented 4 years ago

It should be a N-sized tensor containing H*W.

Ha0Tang commented 4 years ago

Are you mean the size of ns_scr and ns_tgt should be two dimensions, i.e., N x HW? (where H and W means height and width of the input feature, N is the number of batches)

rogerwwww commented 4 years ago

No, it should be [HW, HW, HW, ...] containing N HW

Ha0Tang commented 4 years ago

Thanks, what A_src, A_tgt, G_src, H_src, G_tgt and H_tgt stand for?

Since the features F1 and F2 are directly extracted from images in my task, do you think I need to execute emb1, emb2 = gnn_layer([A_src, emb1], [A_tgt, emb2]) before calculating the affinity?

rogerwwww commented 4 years ago

No, you needn't. A_src, A_tgt, etc are all about the graph connectivity, so you needn't care about them since you are working with images.