search

uber-research / LaneGCN

[ECCV2020 Oral] Learning Lane Graph Representations for Motion Forecasting
https://arxiv.org/abs/2007.13732
Other
496 stars 131 forks source link

Does it really generate graph['node_idcs']? #25

Open HMTJYQS opened 2 years ago

HMTJYQS commented 2 years ago

Hi, Thnanks for release the codes!

I have one question. When I generate graph through preprocess_data.py, I didn't find graph['node_idcs'] was saved in graph this dict, but during training lanegcn.py and executing the MapNet module, I found the following codes in the forward function of MapNet: def forward(self, graph): if ( len(graph["feats"]) == 0 or len(graph["pre"][-1]["u"]) == 0 or len(graph["suc"][-1]["u"]) == 0 ): temp = graph["feats"] return ( temp.new().resize_(0), [temp.new().long().resize_(0) for x in graph["node_idcs"]], temp.new().resize_(0), ) Although there is no error when I running the original module lanegcn.py, Is it because every time the if condition won't be execute or the graph['node_idcs'] was really generated somewhere I didn't find? If this conditional statement is executed, the nodes feature of the output will be empty.

NeoKH commented 2 years ago

@HMTJYQS I'm rewriting the LaneGCN codes recently because of its cumbersome, and I find this unreasonable code too. Is that a BUG? How did you deal with that ? Hope to get your replay.