search

bknyaz / sgg

Train Scene Graph Generation for Visual Genome and GQA in PyTorch >= 1.2 with improved zero and few-shot generalization.
https://arxiv.org/abs/2007.05756
Other
130 stars 20 forks source link

Is there an intention to release the forward code for motifs model? #2

Open galsk87 opened 4 years ago

bknyaz commented 4 years ago

Not in the near future, since it is difficult to make it work in pytorch > 0.3 (just forward pass should be easier though). But I will add it to my todo list.

galsk87 commented 4 years ago

Thanks for the quick reay. How did you get the results reported in the paper?

בתאריך יום ג׳, 9 ביוני 2020, 13:49, מאת Boris Knyazev ‏< notifications@github.com>:

Not in the near future, since it is difficult to make it work in pytorch > 0.3 (just forward pass should be easier though). But I will add it to my todo list.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/bknyaz/sgg/issues/2#issuecomment-641208314, or unsubscribe https://github.com/notifications/unsubscribe-auth/AGZQGMWPF4SWDPSN4EII5QTRVYHTJANCNFSM4NZHQJNA .

bknyaz commented 4 years ago

I used Neural Motif's official implementation with Pytorch 0.3: https://github.com/rowanz/neural-motifs. Just plugged our loss there and followed their steps. In this repo I hoped to make scene graph generation more accessible for users with Pytorch >= 1, but I could only use simple Message Passing and had to change a few functions, that's why results in this repo are somewhat different from the reported in our paper.

galsk87 commented 4 years ago

You can check out code from Tang which you cited for an implementation motis with modern PyTorch.

בתאריך יום ג׳, 9 ביוני 2020, 14:14, מאת Boris Knyazev ‏< notifications@github.com>:

I used Neural Motif's official implementation with Pytorch 0.3: https://github.com/rowanz/neural-motifs. Just plugged our loss there and followed their steps. In this repo I hoped to make scene graph generation more accessible for users with Pytorch >= 1, but I could only use simple Message Passing and had to change a few functions, that's why results in this repo are somewhat different from the reported in our paper.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/bknyaz/sgg/issues/2#issuecomment-641221979, or unsubscribe https://github.com/notifications/unsubscribe-auth/AGZQGMVXZKXM4U3HTHGQNMTRVYKRLANCNFSM4NZHQJNA .